<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Banuba on Medium]]></title>
        <description><![CDATA[Stories by Banuba on Medium]]></description>
        <link>https://medium.com/@banuba?source=rss-e1a005e0267b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*61V4nn-vpv9hcMVe9lMdVA.png</url>
            <title>Stories by Banuba on Medium</title>
            <link>https://medium.com/@banuba?source=rss-e1a005e0267b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 19:57:47 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@banuba/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[6 AR SDK Options Compared for Retail Virtual Try-On (2026)]]></title>
            <link>https://banuba.medium.com/ar-sdk-options-retail-try-on-6691b6bdaeab?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/6691b6bdaeab</guid>
            <category><![CDATA[augmented-reality]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[retail]]></category>
            <category><![CDATA[ecommerce]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 15:34:02 GMT</pubDate>
            <atom:updated>2026-11-17T15:34:02.435Z</atom:updated>
            <content:encoded><![CDATA[<h4>Category coverage, catalog work, hosting, access, published pricing and published results: what a retail team needs before it picks an AR SDK.</h4><p>A retailer adding virtual try-on picks between an AR SDK it embeds itself, a hosted try-on widget, and an enterprise contract with a cosmetics-technology supplier. The three paths differ less in rendering quality than in who produces the product assets, which categories are supported, and whether a price exists in public at all.</p><p>Six vendors, compared on those six terms: Perfect Corp, ModiFace, GlamAR, Banuba, PulpoAR and Auglio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sno5uXiA3fz1jIm31QHqWw.png" /></figure><h3>What each vendor is good for</h3><ul><li>Perfect Corp covers the widest category range in this group, including clothes, shoes, bags and watches, which the five others do not.</li><li>ModiFace has the deepest cosmetics research behind it and L’Oreal distribution, and it is the only vendor here whose technology is already running inside ChatGPT.</li><li>GlamAR publishes a plan ladder with view and SKU limits and reaches in-store screens on its top tier, and it expects you to arrive with 3D models already made.</li><li>Banuba publishes list prices on two self-serve paths and includes product digitization at no extra cost, and it stops at face, head, neck, hands and hair.</li><li>PulpoAR covers cosmetics, skin and hair only, and ships an in-store smart mirror as a product with its own page.</li><li>Auglio has the largest published pre-virtualized catalog, works from a single 2D product photo, and publishes an accuracy figure for pupillary distance that nobody else here matches.</li></ul><h3>How we compared these six</h3><p>Six criteria, each one a decision a retail team makes before launch rather than during development.</p><p><strong>Category coverage.</strong> How many of your product lines the vendor can actually render. A makeup-only engine is a dead end for a store that also sells frames and earrings.</p><p><strong>Who produces the product assets.</strong> Someone has to turn a SKU into something a camera can place on a face. That work is either included, sold separately, or yours.</p><p><strong>Where the try-on runs.</strong> A product page, a native app, an in-store screen, or all three. The answer sets the integration work and often the plan tier.</p><p><strong>Access.</strong> Whether you can sign up and start, or whether you need a sales conversation and an enterprise agreement first. A team can finish a technical evaluation and find there is nothing it is allowed to buy.</p><p><strong>Published pricing.</strong> Whether the cost model is readable in public or only in a quote. Readable pricing lets a retail lead size the pilot without a procurement cycle.</p><p><strong>Published retail outcomes.</strong> What numbers each vendor puts its name to, and who produced them.</p><h3>The comparison table</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GlpBOBKWZ07-6QH1_IpzPA.png" /></figure><p>Category coverage: Perfect Corp wins, because clothes, shoes, bags and watches sit alongside the beauty lines on one vendor’s product list.</p><p>Product assets: Auglio wins, because 500,000+ products are already virtualized and a new one needs a single photograph.</p><p>Where it runs: GlamAR wins, because web, app and in-store appear as named steps on a published plan ladder instead of a scoping call.</p><p>Access: Banuba and Perfect Corp win, because both have a route that starts without a sales conversation.</p><p>Published pricing: Banuba and Perfect Corp are the only two with a number in public, and Banuba publishes on two of its three buying paths against Perfect Corp’s one.</p><p>Published retail outcomes: no vendor wins this one. Every figure in the category is published by the company selling the technology, and three of the six publish no figure at all.</p><h3>Perfect Corp</h3><p><strong>Strengths.</strong> The YouCam product list runs from makeup, hair and nails through eyewear, jewelry and watches to clothes, shoes, bags, scarves and hats, which is the broadest spread here. Its virtual try-on page states that “while 3D models are ideal for categories like shoes, other products such as clothes, jewelry, and makeup can be supported with high-quality 2D images”. Delivery comes as a web SDK, a mobile SDK, an e-commerce web module or an API.</p><p><strong>Limitations.</strong> Enterprise terms are quote-only. The published prices sit on its <a href="https://apps.shopify.com/youcam-makeup-official">Shopify listing</a>, where the two paid tiers are priced per month against a SKU allowance with a 14-day free trial. That listing carries neither pupillary-distance measurement nor frame removal, although <a href="https://www.perfectcorp.com/business/products/virtual-try-on">Perfect Corp’s own try-on page</a> says the eyewear tool “automatically measures pupillary distance”. The capability exists; the self-serve product does not include it.</p><p><strong>Best for</strong> a retailer whose catalog crosses beauty and fashion and wants one vendor across both.</p><p><strong>Skip if</strong> your buying process needs a readable enterprise price before a sales call.</p><p>Perfect Corp’s own figure for makeup try-on is “up to a 200% increase in conversion rates”, attributed on its pages to leading makeup brands with no study, sample or timeframe behind it.</p><h3>ModiFace</h3><p><strong>Strengths.</strong> Founded in 2007 out of research at the University of Toronto and now a L’Oreal company, <a href="https://modiface.com/">ModiFace</a> has the longest research record in cosmetics in this group. Its service list is makeup try-on, hair color try-on, skin analysis and teleconsultation, and its site counter reads “1B+ Virtual Try-On Sessions”.</p><p>Its distribution is the thing to look at. ModiFace powers the Maybelline makeup try-on that now runs inside ChatGPT, announced by L’Oreal and OpenAI in June 2026 and reported as consumer-available in November 2026, reachable by searching for Maybelline in ChatGPT’s Apps section. No other vendor here has shipped a try-on inside a general-purpose assistant. The scope is narrow, one L’Oreal brand and makeup only, and it arrived through a corporate partnership rather than as a capability other retailers can license.</p><p><strong>Limitations.</strong> There is no eyewear, jewelry or accessory try-on in the service list, no pricing page, no plan ladder and no self-serve signup. A legacy nail try-on page still exists but is not in the current navigation. Contact runs through an email address. For an independent retailer this is the hardest vendor in the group to buy.</p><p><strong>Best for</strong> a brand inside the L’Oreal ecosystem, or an enterprise with a partnership route already open.</p><p><strong>Skip if</strong> you sell anything other than makeup, hair color or skincare.</p><h3>GlamAR</h3><p><strong>Strengths.</strong> <a href="https://www.glamar.io/pricing/ar-try-on">GlamAR</a> is the only vendor here that publishes a structured plan ladder: Starter at 10,000 monthly views and 50 SKUs, Growth at 50,000 views and 100 SKUs, Scale at 100,000 views and 500 SKUs. Platform reach climbs with the tier, from web, to web and app, to web, app and in-store, and so does support, from email to priority email and a call. Its FAQ puts onboarding at “2–4 weeks depending on which modules you choose and whether you already have 3D assets ready”, and lists Shopify, WooCommerce, Magento and custom platforms, with SDKs, embed codes or APIs. Category coverage reaches beyond the face into home and luggage.</p><p><strong>Limitations.</strong> The ladder has no prices on it, only a quotation request. More importantly for a budget, the pricing page states that “AR Try-On plans assume you already have AR-ready 3D models”, with 3D model creation priced separately by category and SKU count. There is no free trial, only demo apps and a demo call.</p><p><strong>Best for</strong> a retailer that already owns 3D product assets, or has budget for an asset production project.</p><p><strong>Skip if</strong> your catalog exists as flat product photography and you need try-on live this quarter.</p><p>Worth knowing before you build a business case on it: GlamAR’s own customer story on the jewelry retailer White Cut Diamonds carries no numbers at all. It describes shoppers spending more time exploring designs. If a conversion figure for that account reaches you in a deck, ask which page it came from.</p><h3>Banuba</h3><p><strong>Strengths.</strong> Banuba is one of two vendors here with list prices on a self-serve path, and the only one that publishes them on more than one. Its no-code Easy Virtual Try-On runs $99, $299 and $799 per month for 1,000, 3,500 and 15,000 monthly try-ons, with unlimited products and a 7-day trial on every paid plan. Its Shopify app runs $249, $799 and $1,299 per month for 2,000, 10,000 and 40,000 monthly try-ons, on the same 7-day trial. The structure across all three buying paths is set out in the <a href="https://www.banuba.com/blog/banuba-tint-virtual-try-on-pricing-guide">virtual try-on pricing guide</a>.</p><p>Product digitization is included free on every Easy Virtual Try-On plan, and the dashboard lets a merchant digitize makeup and eyewear SKUs without development work. The self-serve category list is makeup, skincare, eyewear, contact lenses, hair dye, headwear and jewelry. Wigs, goggles, masks and beards sit on the custom integration path.</p><p><strong>Limitations.</strong> Scope stops at the face, head, neck, hands and hair, so clothing, footwear, bags and home goods are categories Banuba does not cover. The Shopify app is narrower than the platform, carrying four categories: makeup, skincare, glasses and hair dye, with no contact lenses. On the self-serve platform the try-on opens as its own page from a link or a QR code rather than inside the product page, which is what the Shopify app and the custom integration are for. White-label branding and a branded domain belong to the custom integration path, which starts with a demo request rather than a signup.</p><p><strong>Best for</strong> a beauty, eyewear or accessory retailer that wants a priced pilot it can start this month and a digitization pipeline it does not have to staff.</p><p><strong>Skip if</strong> you need garment or footwear fitting, or an in-store deployment on a self-serve plan.</p><p>The clearest published result is from Oceane, a Brazilian cosmetics manufacturer and retailer. Its add-to-cart rate on concealer and foundation reached 20.15% in the first month of the pilot, against the 3% industry average the case study uses as its benchmark, and peaked at 32% in July. It is our own case study, on our own technology, with the brand’s pre-pilot rate not published, so read it as a pilot result rather than a controlled test. The arithmetic worth running before a launch is set out in <a href="https://www.linkedin.com/pulse/add-to-cart-math-behind-every-beauty-ar-try-on-launch-banuba-bem0f/">the add-to-cart math behind a beauty AR try-on launch</a>.</p><h3>PulpoAR</h3><p><strong>Strengths.</strong> The solution list stays inside beauty: AR makeup try-on, nail polish try-on, AI shade finder, skin analysis and simulation, hair analysis and color. The Smart Mirror for touchless in-store try-on is a named product with <a href="https://pulpoar.com/smart-mirror/">its own page</a>, which is unusual at this size. Its FAQ states that “we do not store or share any personal images or biometric data”.</p><p><strong>Limitations.</strong> No eyewear, jewelry or accessory try-on, no published pricing, no plan ladder and no self-serve path: every call to action is a demo booking or a contact form. The published outcome figures carry no date, which matters when you are using them to build a business case.</p><p><strong>Best for</strong> a cosmetics brand that wants an online and in-store try-on from one small vendor.</p><p><strong>Skip if</strong> your assortment reaches past makeup, nails, skin and hair.</p><p>On <a href="https://pulpoar.com/">PulpoAR’s homepage</a> three figures sit together: a 12% average conversion rate, 82,000 users adding products to their carts, and a 33% order value uplift. The page labels them that loosely, attaches no client, period or source to any of them, and does not state that the 82,000 is the denominator for the 12%.</p><h3>Auglio</h3><p><strong>Strengths.</strong> <a href="https://auglio.com/">Auglio</a> publishes the largest ready-made catalog in this group, “500.000+ products in our database” available for immediate integration, and states that “all we need to create a Virtual Try-On is a simple 2D photo”. Its eyewear depth is the real differentiator: automatic pupillary-distance measurement, head measurement, face shape detection, photochromic lens visualization, and an Invisible Glasses module that lets a shopper try frames without removing the pair they are wearing.</p><p>It is also the only vendor in this group that puts a tolerance on its PD reading. The <a href="https://auglio.com/en/products/auto-pd-measurement">Auto PD Measurement page</a> states that “our PD measurement achieves an accuracy of ±1.5 mm in 91% of cases”.</p><p><strong>Limitations.</strong> The modules outside eyewear are thin. Cosmetics Try-On and Wig Try-On exist as products and jewelry appears as a solutions industry, but none of them carries the feature depth the eyewear line does. No pricing is published and the route in is a free demo request.</p><p>The PD figure needs reading twice. The same page’s FAQ widens it to “measurements within ±1–2 mm of the actual value”, which is not the same promise, and no methodology, sample size or device list is attached to either number. Auglio’s own featured customer, a digital product manager at an eyewear retailer, describes the automatic PD as “an approximate, is good enough and can be manually adjusted to”. That sentence is worth more than either figure, because it tells you to design the manual override into your checkout.</p><p><strong>Best for</strong> an optical retailer with a large frame catalog and no 3D asset pipeline.</p><p><strong>Skip if</strong> eyewear is one line in a wider beauty assortment.</p><h3>How to choose</h3><p>If your catalog crosses beauty and fashion, Perfect Corp is the only vendor here that covers garments and footwear alongside makeup.</p><p>If you already have AR-ready 3D models, or you need in-store screens on a defined plan rather than a custom project, GlamAR’s ladder is the most legible.</p><p>If you sell frames, Auglio’s half-million-product database removes most of the catalog work before you start, and its published PD tolerance is the only number in this group you can take to a lens lab.</p><p>If you want online and in-store cosmetics try-on from one vendor and do not mind a sales conversation, PulpoAR fits.</p><p>If you are a L’Oreal brand, or you need the deepest skin diagnostics in the group, ModiFace is the technical benchmark.</p><p>If you want to run a priced pilot on a product page next month, across beauty, eyewear and accessories, without producing 3D assets, Banuba is the one with public numbers you can plan against.</p><p>Two patterns show up across all six. The vendors that publish prices publish narrower category lists, and the vendors with the widest coverage ask for a sales call. And every retail outcome figure in this category comes from a company selling try-on, which is a reason to treat the pilot itself as the measurement and to instrument add-to-cart before the widget goes live.</p><p>Our own write-up of <a href="https://banuba.medium.com/7-best-beauty-ar-sdks-compared-2026-195828f7c194">seven beauty AR SDKs compared on engineering criteria</a> covers the developer-side questions this comparison leaves out, and <a href="https://banuba.medium.com/how-to-develop-a-virtual-makeup-app-timing-must-have-features-api-sdk-383e52765ebc">the build timeline for a virtual makeup app</a> covers what a first release usually contains.</p><h3>FAQ</h3><h4><strong>How much does retail virtual try-on cost?</strong></h4><p>Published pricing exists at two vendors. Banuba lists $99 to $799 per month for self-serve Easy Virtual Try-On by monthly try-on volume, and $249 to $1,299 per month for its Shopify app, both on a 7-day trial. Perfect Corp publishes two monthly tiers on its Shopify listing, priced by SKU allowance with a 14-day trial. GlamAR, ModiFace, PulpoAR and Auglio quote on request, so expect a scoping call and a procurement cycle before you see a number from them.</p><h4><strong>Who builds the 3D models and product data?</strong></h4><p>It depends on the vendor, and this is the line item most often missed. GlamAR’s plans assume you arrive with AR-ready 3D models and price their creation separately by category and SKU count. Auglio works from a single 2D photo and already holds over 500,000 virtualized products.</p><p>Banuba includes AI product digitization on every Easy Virtual Try-On plan and lets a merchant digitize makeup and eyewear SKUs from its dashboard. Perfect Corp accepts 2D images for clothes, jewelry and makeup, with 3D reserved for categories like shoes.</p><h4><strong>Can a retailer run virtual try-on without a native app?</strong></h4><p>Yes. Banuba’s Easy Virtual Try-On is a hosted try-on URL or QR code, Perfect Corp ships an e-commerce web module, and GlamAR’s Starter tier is web-only. Native-app try-on is where SDK integration work starts, and Banuba’s documentation and public iOS and Android samples are the route in for that.</p><h4><strong>Which vendors cover eyewear as well as makeup?</strong></h4><p>Perfect Corp, GlamAR, Banuba and Auglio all publish both, though Auglio’s cosmetics module is thin next to its eyewear line. ModiFace and PulpoAR are beauty only. If you sell frames and lipstick from the same store, that is the first criterion to apply, because it halves the list before any other one.</p><h4><strong>Does any vendor publish an accuracy figure for pupillary distance?</strong></h4><p>One does. Auglio states ±1.5 mm in 91% of cases on its Auto PD Measurement page, while its own FAQ widens that to ±1–2 mm. Perfect Corp says its eyewear tool measures PD automatically without putting a tolerance on it. If you sell prescription lenses online, send every vendor the same question in the same words: what share of readings lands within 1 mm, and on which devices was that measured?</p><h4><strong>Does virtual try-on need a Shopify store?</strong></h4><p>No. Banuba’s Shopify app is one of its three buying paths, with Easy Virtual Try-On and a custom integration covering other stacks, and GlamAR lists WooCommerce, Magento and custom platforms alongside Shopify. What happens from camera permission to the cart is covered in our <a href="https://banuba-official.hashnode.dev/virtual-glasses-try-on-sdk-what-runs-between-the-camera-and-the-cart">glasses try-on widget walkthrough</a>.</p><h4><strong>What should a retailer measure in the first month?</strong></h4><p>Add-to-cart rate on the SKUs you enabled against a matched set you did not, return rate on the same SKUs, and try-on sessions per thousand product-page views. Those three answer whether to widen the pilot. Vendor-published conversion figures, Banuba’s included, are a reason to run the test, not a substitute for it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6691b6bdaeab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 Background Removal APIs for Video Apps Compared (2026)]]></title>
            <link>https://banuba.medium.com/background-removal-api-video-apps-125b9ec88646?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/125b9ec88646</guid>
            <category><![CDATA[video-editing]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[augmented-reality]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 15:07:29 GMT</pubDate>
            <atom:updated>2026-11-17T15:20:58.373Z</atom:updated>
            <content:encoded><![CDATA[<h4>In this guide: Apple Vision, ML Kit, MediaPipe, Banuba and VideoBGRemover, scored on five criteria.</h4><p>A background removal API for a video app is a person segmentation model that separates the user from the room in every frame, so the app can swap, blur or clear the backdrop. For live capture the model has to run on the phone or in the browser. For clips that users upload, a cloud service can do the work.</p><p>That split decides almost everything else, so it is the first question to settle. Four of the five options below run on the device. One runs in the cloud and cannot touch a live preview at all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZiUqoefIKCWGKmJiAp1Bag.png" /></figure><h3>How we compared the options</h3><p><strong>Live capture (real time).</strong> The mask keeps pace with the camera preview, or the room flickers through around hair.</p><p><strong>Recorded clips and post-processing.</strong> The same cut-out on every frame of a finished file.</p><p><strong>Platforms.</strong> One integration for every app you ship.</p><p><strong>What you still build.</strong> A mask is half the feature. Somebody still composites the backdrop.</p><p><strong>Access and cost model.</strong> How the option is paid for as usage grows.</p><h3>The 5 options for video apps</h3><h3>1. Apple Vision person segmentation</h3><p>Apple’s option is <a href="https://developer.apple.com/documentation/vision/vngeneratepersonsegmentationrequest">VNGeneratePersonSegmentationRequest</a> in the Vision framework, which produces a matte image for a person it finds in the input image and returns it in a VNPixelBufferObservation. It is available from iOS 15, iPadOS 15, Mac Catalyst 15, macOS 12, tvOS 15, and visionOS 1.</p><p>There are now two symbols for this, and picking the wrong one costs you deployment target. The newer Swift class, <a href="https://developer.apple.com/documentation/vision/generatepersonsegmentationrequest">GeneratePersonSegmentationRequest</a>, has an async perform(on:) and returns a PixelBufferObservation, but it starts at iOS 18 and macOS 15. If your floor is iOS 15, you are on the older class.</p><p>A qualityLevel setting balances accuracy and performance, in Apple&#39;s words. The page publishes no frame rate.</p><p><strong>Strengths:</strong> the mask comes from frameworks your iOS and macOS apps already link, with no new vendor.</p><p><strong>Limitations:</strong> Apple platforms only, and you get a mask, so your team handles compositing.</p><p><strong>Best for:</strong> an iOS-first camera app that wants no new dependency.</p><p><strong>Skip if:</strong> Android or web users matter at launch.</p><h3>2. Google ML Kit Selfie Segmentation</h3><p><a href="https://developers.google.com/ml-kit/vision/selfie-segmentation">ML Kit Selfie Segmentation</a> takes an input image and produces an output mask for static images and live video. Google states that the API is CPU-based and runs in real time on most modern smartphones at 20 FPS or more, which is the only frame rate any of the four mask-only options publishes.</p><p>Read the qualifier with the number. Google names no devices, no OS versions and no input resolution, so “most modern smartphones” is the entire scope that figure carries.</p><p>Two features help on video. It segments multiple people or a single person with no setting change, and during live video it uses output from previous frames to return smoother results. The mask defaults to the input image size, with an option to take the raw model output at 256x256 to cut latency.</p><p><strong>Strengths:</strong> one API on both phones, a published frame rate, multi-person masks.</p><p><strong>Limitations:</strong> Google states the API is offered in beta, not subject to any SLA or deprecation policy, and that changes may break backward compatibility. Android and iOS only, no web, and compositing is yours.</p><p><strong>Best for:</strong> a mobile social app that needs selfie cut-outs on both phones.</p><p><strong>Skip if:</strong> you need an SLA or a web build.</p><h3>3. MediaPipe Image Segmenter</h3><p><a href="https://developers.google.com/edge/mediapipe/solutions/vision/image_segmenter">MediaPipe Image Segmenter</a> divides images into regions by predefined category, with code examples for Web, Android in Kotlin and Java, iOS, Python and Flutter. That is the widest platform list of the mask-only options, and the only one here with a Flutter path.</p><p>Which model you pick decides what you get. The selfie model outputs two categories, background and person. A separate multiclass model covers hair, face, skin, clothing and accessories, and a hair model does hair alone.</p><p>It returns a uint8 category mask where each pixel carries the winning class index, or per-category confidence masks between 0.0 and 1.0. It does not composite a new background. Google&#39;s page names background replacement and blurring as things the selfie model can be used for, which leaves the compositing with you.</p><p>On performance it publishes latency rather than frame rate: the square selfie segmenter is benchmarked at 33.46 ms on CPU and 35.15 ms on GPU, on a Pixel 6. One device, one model, so treat it as a starting point rather than a range.</p><p>The LIVE_STREAM mode is asynchronous and automatically drops incoming frames when busy so results stay current, while VIDEO mode processes every frame in order and never drops any. The web path has no LIVE_STREAM option and pairs VIDEO with requestAnimationFrame instead, which matters if your browser build is the one that has to keep up.</p><p><strong>Strengths:</strong> six platform targets under an Apache 2.0 license, with a published latency figure on a named device.</p><p><strong>Limitations:</strong> masks only. You own compositing, frame pacing and model updates.</p><p><strong>Best for:</strong> a team with computer vision engineers that wants one open codebase.</p><p><strong>Skip if:</strong> nobody wants to own the rendering.</p><h3>4. Banuba Face AR SDK (Virtual Background API)</h3><p>Banuba Face AR SDK is a real-time, on-device face tracking and AR effects SDK, and its <a href="https://www.banuba.com/background-changer-api">Virtual Background API</a> is listed as available in real time and in post-processing. Unlike the three above, it composites the result rather than handing back a mask: Background.texture() sets an image or a video file as the backdrop, and Background.blur() takes a radius from 0 to 1.</p><p>The figure to plan against is the background separation row of our technical specification, not the headline one. It lists 25 FPS on low-end Android, and 30 FPS on high-end Android and on mid and high iOS devices. Our own docs also note that performance can be lower on mid-end and low-end devices because of neural network usage, which is the same warning in the other direction.</p><p>Face AR SDK ships for Android, iOS, Web, macOS, Windows, Unity, React Native and Flutter. For the web build specifically, npm pack @banuba/webar@1.18.6 with only the background module brings the SIMD path to about 4.8 MB gzipped, which you can reproduce with that one command rather than taking our word for it.</p><p>Our <a href="https://banuba-official.hashnode.dev/background-removal-api-in-5-steps-replace-blur-or-clear-a-webcam-backdrop">five-step guide to replacing, blurring or clearing a webcam backdrop</a> shows those calls in a browser. The <a href="https://www.banuba.com/blog/over-20m-downloads-videoshop-success-story">Videoshop case study</a> names advanced green screen features among the results, and quotes the team: “Banuba’s virtual background technology is one of the best on the market.”</p><p><strong>Strengths:</strong> a composited backdrop, frame rates broken out by device tier, and mobile, web and desktop from one SDK.</p><p><strong>Limitations:</strong> a commercial license, with custom MAU-based pricing and no public price list.</p><p><strong>Best for:</strong> a creator, social video or conferencing app that needs the swap in the live preview.</p><p><strong>Skip if:</strong> your users only upload finished files.</p><h3>5. VideoBGRemover</h3><p><a href="https://videobgremover.com/">VideoBGRemover</a> is a cloud video background remover: you upload files or pass public URLs, and its Python and Node.js SDKs return either a mask video or a composed result with a custom color, image or video behind the subject. Processing happens on its servers, so no local GPU is involved.</p><p>It accepts MP4, MOV and WebM, up to two hours per video. Its pricing page offers up to 4K, while the same page’s FAQ says processing runs at HD 1080p and that 4K results may vary, so confirm the resolution you actually need before you commit a workflow to it.</p><p>Budget the wait rather than the upload. Its own documentation puts processing at typically 4 to 5 minutes per minute of video, which makes a five-minute clip a twenty-minute job, and lists frame rate, resolution, model and current capacity as the things that move it.</p><p>Pricing runs from $4.80 a minute at the entry tier down to $2.50 a minute above $1,000 of cumulative purchase, bought as credit packs that expire one year after purchase. Frame rate multiplies the cost: 30 FPS and under counts once, 31 to 60 FPS counts twice, and above 120 FPS counts four times. Treat those as list prices rather than a quote.</p><p><strong>Strengths:</strong> nothing runs on the user’s device, and long files come back composited.</p><p><strong>Limitations:</strong> it works on uploaded files, so it cannot cut out the room in a live camera preview. The bill grows with every minute processed, and faster footage costs more per minute than slower footage.</p><p><strong>Best for:</strong> an editor where users upload finished clips and wait for the export.</p><p><strong>Skip if:</strong> the effect must appear while recording.</p><h3>Which background removal API wins each criterion?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4n2IaKsRH8QOS3dijHr9Ug.png" /></figure><ul><li><strong>Apple Vision</strong> is person segmentation that publishes no live-capture figure, leaves you to build the frame loop for recorded clips, runs on Apple operating systems only, hands back a mask so compositing and the backdrop picker are yours, and arrives inside the Vision framework at no extra cost.</li><li><strong>ML Kit Selfie Segmentation</strong> publishes 20 FPS or more as Google’s own claim, handles images and video, covers Android and iOS, leaves compositing and the picker with you, and is free but in beta with no SLA.</li><li><strong>MediaPipe Image Segmenter</strong> publishes 33.46 ms per frame on a Pixel 6 rather than a frame rate, offers IMAGE and VIDEO modes, reaches Web, Android, iOS, Python and Flutter, leaves compositing and the picker with you, and is open source under Apache 2.0.</li><li><strong>Banuba Face AR SDK</strong> lists 25 to 30 FPS by device tier for background separation, works in real time and in post-processing, ships for eight targets including desktop, returns the composite so only the picker and the backdrops are yours, and is a commercial license with custom MAU-based pricing and a 14-day trial.</li><li><strong>VideoBGRemover</strong> offers no live capture at all, takes uploaded files up to two hours, is a cloud API with Python and Node.js SDKs, returns the composite so you build the upload flow and manage the wait, and bills $2.50 to $4.80 a minute on credits that expire after a year.</li></ul><p><strong>Live capture: Banuba and ML Kit are the only two that publish a frame rate.</strong> Banuba lists 25 to 30 FPS by device tier, Google claims 20 FPS or more without naming a device. MediaPipe publishes a millisecond latency on one phone instead, and Apple publishes nothing.</p><p><strong>Recorded clips: VideoBGRemover wins</strong>, with uploaded files up to two hours and the composite returned.</p><p><strong>Platforms: Banuba wins</strong>, with eight targets from one SDK, desktop included.</p><p><strong>What you still build: Banuba and VideoBGRemover tie</strong>, as both return the composite rather than a mask.</p><p><strong>Access and cost model: MediaPipe wins</strong> as open source under Apache 2.0, with Apple Vision tying for Apple-only apps.</p><h3>How do you choose a background removal API for a video app?</h3><ul><li>If you are building an iOS-only camera app, use Apple Vision, because the person mask ships in a framework your app already links.</li><li>If you are building a social app on Android and iOS with your own compositing, use ML Kit, because one API covers both and it publishes a frame rate.</li><li>If you need web, mobile and a Python job from one open codebase, use MediaPipe, because it is Apache 2.0 and reaches six targets including Flutter.</li><li>If you are building a creator or conferencing app that needs the swap live on mobile, web and desktop, use Banuba, because it composites the backdrop. For the calling layer under a conferencing app, see our <a href="https://banuba.medium.com/12-best-video-conferencing-apis-and-sdks-in-2026-tested-and-compared-768d04958577">comparison of video conferencing APIs and SDKs</a>.</li><li>If your users upload finished clips and are willing to wait, use VideoBGRemover, because it handles long files in the cloud without touching the device.</li></ul><h3>Why Banuba</h3><p>Banuba Face AR SDK removes, blurs or replaces the background on the device, live or in post-processing, next to face tracking and AR effects. It is not a cloud batch service for uploaded files, and VideoBGRemover fits that job better.</p><p>It is a commercial license, and three inputs set the quote: the platforms you ship on, the feature categories you activate, and your MAU. A 14-day free trial covers every feature on every supported platform.</p><p>If a mask is enough, you have cheaper options and we would rather you knew that. Apple Vision ships inside Apple’s own framework, and MediaPipe is open source under Apache 2.0.</p><h3>Live preview needs an on-device SDK, uploaded clips can go to a cloud API</h3><p>A swap in the camera preview needs a model on the phone or in the browser, and Apple Vision, ML Kit, MediaPipe and Banuba differ mainly in how much compositing you own after the mask arrives.</p><p>Uploaded files fit a remove background API in the cloud such as VideoBGRemover, where the trade is the opposite: no device cost, but a per-minute bill and a wait measured in multiples of the clip’s length.</p><h3>FAQ</h3><h4>How does Banuba’s license compare with MediaPipe and ML Kit?</h4><p>Banuba Face AR SDK is a commercial license with custom MAU-based pricing and a 14-day free trial. MediaPipe is open source under Apache 2.0. ML Kit is free but Google states it is in beta with no SLA or deprecation policy.</p><h4>What is the best background removal API for live video?</h4><p>Banuba Face AR SDK lists background separation at 25 to 30 FPS by device tier and returns the composited frame. ML Kit publishes 20 FPS or more and returns a mask. Those are the only two published frame rates among the five, so for anything else you measure it yourself on your lowest supported device.</p><h4>How large is the Banuba Web SDK with only the background module?</h4><p>About 4.8 MB gzipped on the SIMD path, measured with npm pack @banuba/webar@1.18.6 for the web build with the background module alone. Run the same command to check it against the version you plan to ship.</p><h4>Does Banuba background removal run in a web browser?</h4><p>Yes. Banuba Web SDK supports any browser with WebGL 2.0 or higher. MediaPipe also runs on the web, though without its frame-dropping stream mode there. The Apple Vision and ML Kit pages list no web build.</p><h4>Can a cloud background removal API work on a live camera preview?</h4><p>No. A cloud service processes an uploaded file, so the round trip rules out a live preview. VideoBGRemover’s own documentation puts processing at typically 4 to 5 minutes per minute of video, which is the scale of delay involved. Live capture needs the model on the device or in the browser.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=125b9ec88646" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[6 Video Editing API Options Compared by Where the Render Runs (2026)]]></title>
            <link>https://banuba.medium.com/video-editing-api-options-compared-79235c4e54fe?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/79235c4e54fe</guid>
            <category><![CDATA[video-streaming-service]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[video-editing]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 13:47:26 GMT</pubDate>
            <atom:updated>2026-11-03T13:47:26.427Z</atom:updated>
            <content:encoded><![CDATA[<h4>A video editing API either renders on a server you call, renders on the phone in your user’s hands, or hands you raw frames and expects you to build the editor yourself.</h4><p>Those three are not substitutes, so picking the wrong one burns a release cycle before anyone notices.</p><p>In this guide we group six options by that single question and name what each one gives up.</p><p>Banuba Video Editor SDK is a production-ready mobile video editing engine for iOS, Android, Flutter and React Native, and Banuba’s Video Editing API is one of the six below. We published this comparison, so the five other entries are described from their own documentation rather than from any test we ran against them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xMv2XZ_qlPusIsX4vu8y6w.png" /></figure><h3>What does “video editing API” mean to three different buyers?</h3><p>The phrase covers three products that share almost nothing operationally.</p><p>A cloud rendering service takes a description of an edit, usually JSON, renders it on the vendor’s infrastructure, and returns a file. There is no camera and no live preview, and you pay per render. It suits video produced with no user present.</p><p>A client-side editing engine ships inside your app binary, drives the camera, applies effects to the live preview, and exports on the user’s own device. There is no render farm, so there is no per-render fee, and the media does not leave the handset.</p><p>A media framework gives you decoders, encoders and frames, and no editor. The timeline, the preview and the export pipeline are all code you write.</p><h3>Renders on the vendor’s servers</h3><h4><strong>Shotstack</strong></h4><p>A cloud render platform with a REST API and JSON edit templates, plus server-side SDKs for Node.js, Python, PHP and Ruby. Shotstack also sells a separate white-label editor SDK for embedding an editor in a web product, so treating it as server-only is wrong. Sign-up is self-serve, with a free tier to start.</p><h3>Renders on your user’s device</h3><h4><strong>Banuba Video Editing API</strong></h4><p>The same engine as our Video Editor SDK with the interface removed, so your screens and user flow stay yours. Recording, editing and export all run on the device, and our <a href="https://www.banuba.com/video-editing-api">video editing API</a> page states that user media is processed locally and is not uploaded to Banuba.</p><p>That page puts integration at 1 to 2 days on a supported platform and says the API runs smoothly on 90% of mobile phones. Platform floors are iOS 15 with Swift 5.9 and Xcode 26.0, and Android 8.0 with Kotlin 2.1 or Java 17, the Camera 2 API and OpenGL ES 3.0, rising to 3.1 when neural networks run on the GPU.</p><h3>Hands you frames, not an editor</h3><h4><strong>AVFoundation</strong></h4><p>Apple’s own media framework, and the foundation under most creative tools on iOS and macOS. It reaches formats nothing else on this list touches, including HDR, Dolby Vision and ProRes. It is free with the platform and needs an Apple Developer Program membership to ship. It exists only on Apple hardware, and concepts like AVMutableComposition make it specialist work.</p><h4><strong>Jetpack Media3 Transformer</strong></h4><p>Google’s answer to Android encoder fragmentation, open source and part of the standard Jetpack suite. It prioritizes transmuxing, so trims and rotations skip a full re-encode and finish fast. No sign-up and no license fee. Android only, and it carries no AI effects layer.</p><h4><strong>FFmpeg</strong></h4><p>An open-source command-line engine that decodes, encodes, muxes and filters effectively any audio or video format, on servers and on mobile alike. Its format coverage is why our own comparison names it the right tool for server-side batch work. There is no UI and no sign-up, and commercial builds need care over how codecs are compiled.</p><h4><strong>WebCodecs</strong></h4><p>A standardized browser API, free, that exposes the device’s hardware encoders and decoders to web code. It brought near-native frame handling into the browser. It returns raw frames and nothing more, which leaves the trimming, the stitching and the timeline to you.</p><h3>Which option wins which criterion?</h3><p>Each winner below is that vendor’s own published positioning, read in November 2026, not a measurement we ran against them. Vendor documentation changes without notice, so any figure that decides a build should come from the vendor in writing before you commit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WPHHb4aOn0x3_ehi8lDoSg.png" /></figure><p>Criterion by criterion, since the table above is an image:</p><ul><li><strong>Where the render runs.</strong> Shotstack renders on the vendor’s servers, Banuba renders on the user’s device. This is a category choice rather than a quality judgment, and answering it first removes most of the rest of the table.</li><li><strong>Bulk video with no user present:</strong> Shotstack, because it renders from a JSON description on demand and no client-side engine does that.</li><li><strong>Server-side batch transcoding:</strong> FFmpeg, on format and codec coverage. No client-side SDK replaces it for this work, ours included.</li><li><strong>In-app creator experience across platforms:</strong> Banuba, because one integration covers iOS, Android, Flutter and React Native instead of two native media stacks.</li><li><strong>iOS-only precision formats:</strong> AVFoundation, because HDR, Dolby Vision and ProRes are reachable there and not through a cross-platform engine.</li><li><strong>Fast trims on Android at zero license cost:</strong> Jetpack Media3, because transmuxing avoids re-encoding and the library ships with Android.</li><li><strong>Editing inside a browser:</strong> WebCodecs, and it is not close, because our editor does not run on the web at all.</li><li><strong>Published performance figures:</strong> Banuba, as the only entry here that publishes export time and preview frame rate by device tier, measured in its own lab.</li><li><strong>Availability:</strong> FFmpeg, Jetpack Media3 and WebCodecs are free and open with no sign-up. Shotstack is self-serve with a free tier. Banuba needs a trial token and then a contract. AVFoundation needs an Apple Developer Program membership.</li></ul><h3>What performance figures does anyone actually publish?</h3><p>Almost none, which is the honest state of this category. Four of the six are frameworks with no editor, so there is nothing to benchmark until you have built something, and a cloud render service’s speed is a function of your template and their queue rather than a number they can print.</p><p>We publish ours, and they are our own lab rather than an independent test. Our <a href="https://www.banuba.com/blog/ve-sdk-performance-benchmark-compared">video editor SDK performance benchmark</a> gives an export of about 3 seconds for an 8-second near-4K clip, 30 or more frames per second in preview on mid-range devices with AR effects, segmentation and beautification enabled, and a tier split where mid-range devices record and export at 720p30 while high-end devices reach 1080p30.</p><p>Treat those as a starting point and not a guarantee. The page names a tested configuration rather than specific handsets, so the number that matters to you is the one you measure on the slowest phone in your own support matrix.</p><h3>How much does a video editing API cost to license?</h3><p>FFmpeg, Jetpack Media3 and WebCodecs cost nothing to use and bill you in specialist hours instead, because you own whatever editor you build on them. AVFoundation is the same trade with an Apple Developer Program membership attached.</p><p>Shotstack publishes its pricing on its own pricing page, which is where that figure should come from.</p><p>Our own licensing has two inputs: monthly active users, and the platforms you ship on. It is a yearly per-app license covering every platform inside one agreement, with no per-developer seat fee, and a 14-day free trial with the full feature set and no credit card.</p><p>There is no free production tier, and SDK customization, white-label branding, integration assistance and custom AR assets are quoted on top rather than included. The <a href="https://www.banuba.com/blog/banuba-video-editor-sdk-pricing-guide">Video Editor SDK pricing guide</a> sets out how the bands work.</p><h3>Why teams pick Banuba, and where we are the wrong answer</h3><p>Banuba’s Video Editing API is a client-side mobile editing engine. It is not a cloud rendering service, it does not run in a browser, and it does not do server-side batch work with no user present. Those are separate categories, and FFmpeg or a hosted render API serves them properly.</p><p>Inside its own category the argument is platform reach and elapsed time. One integration covers iOS, Android, Flutter and React Native, which is the reason teams choose it over writing AVFoundation and Jetpack Media3 code twice and maintaining both.</p><p>Weat, a US social commerce platform, built its video-native app on our engine and <a href="https://banuba.medium.com/weat-video-editing-api-case-b3aa13cec7ca">reported saving 50% of its product development time</a> against building the editor in-house. That is the customer’s own measurement of their own project, not a controlled test, and your number will differ.</p><p>One more distinction is worth knowing before you ask us for the wrong product. The API exposes the engine without an interface, while the Video Editor SDK ships a finished editor instead. The SDK cannot reorder its screens after the entry point, which is the specific reason the API exists. Our <a href="https://medium.com/@banuba/7-best-mobile-video-editing-sdks-feature-by-feature-comparison-1b3519c8592">7 best mobile video editing SDKs comparison</a> covers that embeddable-editor category feature by feature.</p><h3>Choosing in one pass</h3><p>Answer where the render has to happen and five of the six options disappear.</p><p>If a server renders video with no user present, you want a cloud API or FFmpeg, and the deciding factor is whether you would rather operate the infrastructure or pay someone else to. If a user records and edits inside your app on more than one platform, you want a client-side engine, and the deciding factor is how much of the editor you intend to write.</p><p>If the editing happens in a browser, WebCodecs is the route and you are building the editor on top of it. If you ship on Apple only and need professional formats, AVFoundation repays the specialist hiring.</p><p>Getting the vendor wrong inside the right category costs a feature you can add later. Getting the category wrong means rewriting the integration, so the render location belongs at the start of the evaluation rather than the end of it.</p><h3>FAQ</h3><h4>Does a video editing API like Banuba’s run on my server or inside my app?</h4><p>Both kinds exist and they are not interchangeable. A cloud rendering API runs on the vendor’s servers and returns a finished file, which suits automated production. Banuba’s Video Editing API runs inside your iOS, Android, Flutter or React Native app, drives the camera and live preview, and keeps user media on the device.</p><h4>Can Banuba replace FFmpeg on our backend?</h4><p>No. Banuba’s engine is client-side, so it does not transcode files on a server with no user present. FFmpeg is the right tool for that work and we say so on our own comparison page.</p><h4>Does Banuba’s Video Editing API work on the web?</h4><p>Not for the editor. Our editor supports iOS, Android, Flutter and React Native, and web is not a supported platform for it. A browser-based editor is a WebCodecs or WebAssembly project.</p><h4>How fast does Banuba’s Video Editing API export a video?</h4><p>Our own benchmark puts an 8-second near-4K export at about 3 seconds, with preview holding 30 or more frames per second on mid-range devices with AR effects, segmentation and beautification enabled. Those figures come from our lab on a stated test configuration rather than from an independent test, so measure the export on the slowest phone you intend to support.</p><h4>Can an AI coding assistant integrate Banuba’s Video Editing API?</h4><p>Partly. We publish <a href="https://github.com/Banuba/ai-skills">Agent Skills</a> that Claude Code, Codex and Qwen Code read directly instead of inferring the API surface, and the human engineer still owns the architecture and the release.</p><h4>What are the minimum platform versions for Banuba’s Video Editing API?</h4><p>iOS 15 with Swift 5.9 and Xcode 26.0, and Android 8.0 with Kotlin 2.1 or Java 17, the Camera 2 API and OpenGL ES 3.0, which becomes 3.1 when neural networks run on the GPU. The current per-platform requirements are in <a href="https://docs.banuba.com/ve-pe-sdk/docs/ios/requirements/">our documentation</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=79235c4e54fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[7 Face AR SDKs Compared for Creator Apps in 2026]]></title>
            <link>https://banuba.medium.com/face-ar-sdk-comparison-creator-apps-3a9fe7b0772b?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/3a9fe7b0772b</guid>
            <category><![CDATA[live-streaming]]></category>
            <category><![CDATA[sdk]]></category>
            <category><![CDATA[augmented-reality]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 16:47:16 GMT</pubDate>
            <atom:updated>2026-10-28T16:47:16.279Z</atom:updated>
            <content:encoded><![CDATA[<h4>Choosing a face AR SDK for creator apps is a ranking exercise: platform reach, documented tracking output, runtime cost, what you pay, who has to approve you, and the license you sign.</h4><p>In this guide we set out seven selection criteria, put seven providers through them, and name the winner of each.</p><p>One disclosure first. We build one of the seven products below, which is why our criteria are stated before the list and why each product gets the same read, including the criteria where other providers beat us.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F4DPOKdjFeimGCJoKPab8w.png" /></figure><h3>How should you evaluate a face AR SDK for a creator app?</h3><p>Rank these seven for your own app before you open a single product page. Every option here is strong on something, so the ranking makes the choice.</p><ul><li><strong>Platform coverage from one integration<br></strong>How many of your targets (iOS, Android, Web, Unity, Flutter, React Native, desktop) come out of one engineering effort, and whether each plugin is maintained by the provider or by a community.</li><li><strong>Documented tracking output<br></strong>Landmark or keypoint count, mesh detail, head-angle range, and how many faces per frame the provider puts in writing. Co-streams, duets and group rooms make the faces-per-frame number load-bearing.</li><li><strong>Published runtime cost on the device<br></strong>Frame rate on real hardware, binary size added to your app, memory. A creator audience sits on mid-tier Android, so a figure measured on a flagship answers a different question.</li><li><strong>Cost and the access gate<br></strong>What it costs, and separately, how long it takes to get your hands on it: open sign-up, an approval queue, a trial application, or nothing at all.</li><li><strong>License terms<br></strong>Open source, free, and free tier are three different things. Only an open-source license lets you read the terms, fork the code and ship it without asking.</li><li><strong>Where the effects come from<br></strong>A tracking engine with no effect supply means your team authors every mask, while a ready catalog shortens the first release.</li><li><strong>Makeup and hair rendering<br></strong>Beauty effects carry a large share of camera opens in selfie-led products, and support ranges from shipped to beta to absent.</li></ul><p>Every competitor figure below is the provider’s own published statement, read from their documentation and product pages between July and October 2026. Any price quoted is approximate and is not a quote. Provider pages change, so a number that decides your build should come from the provider in writing before you commit.</p><h3>The seven options side by side</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MT5EGAev4VjRIJwyeK425A.png" /></figure><p>The same seven rows in prose, one provider per line:</p><h4><strong>DeepAR</strong></h4><p>iOS, Android, macOS and Web, with Windows listed as coming soon. It does not publish a maximum face count, but it does publish a 60-degree angle limit, a 150 to 180 cm range, a 30 FPS cap on iOS and Web, and binary sizes of 34.6 MB on iOS and 32.3 MB on Android. Self-serve sign-up, a public MAU price table, a commercial license, its own effects, and a Beauty API still in beta.</p><h4><strong>Snap Camera Kit</strong></h4><p>iOS, Android and Web. No tracking numbers and no runtime figures are published, and no price appears on any Camera Kit page. Access needs a developer account and an access request, the first release is reviewed by Snap, and the license carries Camera Kit obligations around Creative Kit and Snapchat sharing. Effects are the existing Snap Lens catalog.</p><h4><strong>Banuba Face AR SDK</strong></h4><p>Eight targets from one integration. 68 anchor points, a mesh of up to 3,308 vertices driven by 37 morphs, five faces as the headline and three recommended on mobile. Performance published per device tier. A 14-day trial on request, a commercial license, no public price list, and over 1,000 Asset Store filters plus Banuba Studio.</p><h4><strong>BytePlus Effects</strong></h4><p>iOS and Android, with face detection also on Windows and macOS and no browser target. Keypoint models at 106, 240 and 280 points, and 10 faces plus 10 portraits per frame. Detection under 2 ms on an iPhone 7 and under 4 MB of memory on an OPPO R11. Sales-led pricing, an offline license file bound to bundle ID and expiry, and its own effects with age, gender and expression attributes.</p><h4><strong>Faceunity</strong></h4><p>iOS, Android, PC, Mac, Unity, Flutter and Electron, with no browser target. No tracking output and no runtime figures published. A free trial by application, sales-led pricing, a commercial license, and the most specific makeup support of any page here: full-face styles plus hairstyle and color adjustment.</p><h4><strong>MediaPipe</strong></h4><p>Implementation guides for Android, Python and Web, with iOS unconfirmed. 478 3D landmarks, with smoothing applied only at a face count of one. No runtime figures, no price, no account needed, and an Apache 2.0 license. Landmarks only, so every effect is yours to build.</p><h4><strong>Google ML Kit</strong></h4><p>Face Detection on Android and 64-bit iOS, Face Mesh on Android alone, no web SDK. Contours for the most prominent face only. About 6.9 MB bundled or 800 KB unbundled for Face Detection, about 6.4 MB for Face Mesh. Free, no account, closed source, detection only.</p><p><strong>Banuba wins platform coverage:</strong> eight targets from one integration, with the Flutter and React Native plugins built and maintained by us, so no community wrapper sits in the chain.</p><p><strong>BytePlus wins documented tracking capacity:</strong> it states 10 faces and 10 portraits per frame, with keypoint models at 106, 240 and 280 points, against the three simultaneous faces we recommend on mobile.</p><blockquote>Banuba wins published runtime evidence, and this is the one criterion where we can show the receipt.</blockquote><p>Our <a href="https://docs.banuba.com/far-sdk/tutorials/capabilities/technical_specification">technical specification</a> breaks performance down by device tier, measured in our own lab: single-face tracking at 25 FPS on the Android Low tier and 30 FPS on Android High and both iOS tiers, 80 degrees of head angle on every tier, and maximum face tracking distance of 170 cm on Android Low, 180 cm on Android High and 230 cm on both iOS tiers.</p><p>No other page we read breaks performance down by device at all. Quote any of those figures with the name of the row attached, because the same document gives 29 FPS for Face AR effects on Android High where the single-face tracking row gives 30. DeepAR takes the binary-size half of this criterion, publishing per-platform figures that we do not.</p><p><strong>Google ML Kit wins on cost:</strong> Google documents the ML Kit face APIs as offered at no cost with no paid tier, and DeepAR is the only paid provider here that publishes a price table at all.</p><p><strong>MediaPipe wins on license:</strong> the repository’s own LICENSE file is Apache 2.0, so the terms are readable and the code is forkable, which no commercial option on this list allows.</p><p><strong>Snap Camera Kit wins effect supply:</strong> Camera Kit brings existing Snap Lenses into your app, so a team can ship recognizable effects before it has authored any.</p><p><strong>Faceunity wins documented makeup coverage:</strong> its product page names full-face makeup styles with hairstyle and color adjustment as shipped, ahead of DeepAR’s beta Beauty API and ahead of what the other pages we read state.</p><h3>Which of these can you sign up for today, and which need approval?</h3><p>Access is a schedule risk, so treat it like a technical criterion. MediaPipe and Google ML Kit need nothing: no account, no application, no approval. DeepAR is self-serve, and a Banuba trial is a 14-day request.</p><p>Three of the seven sit behind a human. Faceunity grants its trial on application, and BytePlus routes through sales while stating that its trial license must not be used for an App Store release, with the risk falling on the customer.</p><p>Snap is the heaviest gate, and it is contractual as well as procedural. It issues an API token per application after an access request, and a first release using Camera Kit must be approved by the Snap Kit team. Snap’s terms state that material changes require re-submission to and re-review by Snap for approval, and that Snap may re-review any use of Camera Kit at any time.</p><p>The same terms oblige you to integrate Creative Kit for sharing content featuring Lenses, to list Snapchat as one of the first several sharing options outside your app, and to use commercially reasonable efforts to make the majority of your Lenses available on Snapchat. If your retention model depends on content staying inside your own app, read those obligations before the build starts.</p><h3>DeepAR</h3><p>DeepAR, from I Love Ice Cream Ltd., is the most transparently documented option here. Its docs list iOS, macOS, Android and Web as the supported platforms, and the marketing site’s Unity claim is contradicted by that list, so plan against the documented four.</p><p>Its <a href="https://docs.deepar.ai/deepar-sdk/pricing">public MAU pricing table</a> runs free to 10 monthly active users, $25 a month to 1,000, $100 to 5,000, $500 to 30,000 and $1,000 to 100,000, with anything above that going to sales. One condition on that table is worth modeling before you compare it with anyone else’s user-based price: DeepAR states that a user who switches devices or browsers may be counted separately for each device and browser combination in the same month.</p><p>The published limits are equally specific. Emotion detection is unavailable on Web, hair segmentation is iOS only, body tracking is in closed beta, and retail try-on tracking sits in ShopAR rather than the base SDK. Maximum simultaneous faces is the one figure DeepAR leaves out. If DeepAR is your other finalist, <a href="https://banuba.medium.com/banuba-face-mask-sdk-vs-deepar-which-one-ships-to-production-in-2026-eab4c54c5abf">our production-readiness comparison of DeepAR and Banuba</a> covers the same ground at depth.</p><h3>Snap Camera Kit</h3><p><a href="https://www.snap.com/terms/snap-camera-kit">Snap Camera Kit</a> puts Snap Lenses inside your own iOS, Android or web app, with Android requiring OpenGL ES 3.0 and Web running on Chrome 95+ and Safari 16+ while Edge and Firefox are under evaluation. The staging token watermarks the screen and applies lower rate limits, and the production token removes both.</p><p>Two integration details decide feasibility early. The web runtime downloads a WebAssembly Lens rendering engine from a Snap-managed CDN, so your Content Security Policy must allow the Snap AR domain plus blob: and wasm-unsafe-eval, or Camera Kit Web will not initialize.</p><p>Attribution is the second. It is optional on mobile and web even when you monetize, but required for commercial public displays, which Snap defines to include retail AR mirrors and kiosks. No price or usage rate appears on any Camera Kit page, and Unity support is undocumented.</p><h3>Banuba Face AR SDK</h3><p>Our SDK processes every frame on the device, with no cloud round-trip, and works offline. The tracking figures on the <a href="https://www.banuba.com/facear-sdk/face-filters">product page</a> are 68 facial anchor points, a 3D mesh of up to 3,308 vertices driven by 37 tracked morphs, and tracking under up to 70% occlusion.</p><p>The performance figures sit in a different document, and that separation is worth knowing. Frame rate, head angle and tracking distance are published per device tier in the technical specification, quoted in full above. The mobile build is about 15 MB depending on the features you enable. If a number matters to your build, take the technical specification first and treat the product page as the one to confirm with us.</p><p>On simultaneous faces our own guidance is split, so take the conservative number: five is the headline capability and three is what we recommend on mobile, because past that the on-device compute budget starts costing frames. Pricing is a flat annual fee with unlimited usage or an active-user model that costs less early and grows with the user base, quoted through sales rather than published.</p><h3>BytePlus Effects</h3><p><a href="https://www.byteplus.com/en/product/effects">BytePlus Effects</a> leads on raw face capacity, at 10 faces and 10 portraits per frame, with keypoint models in 106, 240 and 280 point configurations. BytePlus also publishes narrow but useful speed figures: detection under 2 ms on an iPhone 7 and under 4 MB of memory on an OPPO R11.</p><p>The constraints are platform shape and licensing mechanics. Browser support is absent from the technical requirements, and 3D face mesh and skin segmentation are documented for Android and iOS only.</p><p>The license file is the detail that catches teams late. It validates both your ApplicationID or BundleID and an expiry date, so a mismatch fails the SDK outright. Pricing is sales-led with nothing published.</p><h3>Faceunity</h3><p>Faceunity, from Hangzhou Xiangxin Technology, sells the Nama SDK for face AR and beautification across iOS, Android, PC, Mac, Unity, Flutter and Electron, with browser and HTML5 absent from the list. Its makeup support is the most specific of any page we read: full-face makeup styles plus hairstyle and color adjustment.</p><p>It also publishes materially less than everyone else here, which shapes how you can evaluate it. No per-device performance figures, no SDK size, no keypoint count and no license terms appear on the <a href="https://www.faceunity.com/en/effects.html">Faceunity product page</a>, so scoring it on the first three criteria above takes a trial application and your own benchmark.</p><h3>MediaPipe</h3><p>MediaPipe’s Face Landmarker is the open-source entry, Apache 2.0 by the repository’s own LICENSE file, and it estimates 478 three-dimensional face landmarks using a BlazeFace short-range model. Implementation guides cover Android, Python and Web, iOS support is unconfirmed in the docs, and Google labels the release an early preview.</p><p>Three details change the engineering estimate. Landmark smoothing is applied only when the face count is set to 1, blendshape and transformation-matrix outputs are off by default, and the Tasks APIs send performance and utilization metrics to Google, with consent obligations left to you.</p><p>Support is a Slack community, GitHub issues and Stack Overflow, with no SLA and no commercial backstop, which is the tradeoff <a href="https://banuba.medium.com/banuba-face-ar-sdk-vs-opencv-a-technical-comparison-1ba68df81c3a">our comparison of OpenCV and Banuba face detection</a> works through in the same terms. The <a href="https://github.com/google-ai-edge/mediapipe">MediaPipe repository</a> is where you confirm the license yourself.</p><h3>Google ML Kit</h3><p>Google documents the <a href="https://developers.google.com/ml-kit/guides">ML Kit APIs</a> as offered at no cost, with no pricing page and no paid tier. It is a closed-source SDK that is free, a different proposition from MediaPipe’s Apache 2.0 license, and the distinction bites the moment you need to change behavior instead of configuring it.</p><p>Two separate face products exist. <a href="https://developers.google.com/ml-kit/vision/face-detection/android">Face Detection</a> is generally available on Android and 64-bit iOS, and adds about 6.9 MB bundled or about 800 KB unbundled. <a href="https://developers.google.com/ml-kit/vision/face-mesh-detection">Face Mesh Detection</a> is Android-only, bundled only, about 6.4 MB, and Google states that it is offered in beta, is not subject to any SLA or deprecation policy, and may change in ways that break backward compatibility.</p><p>Platform reach is the narrowest of the seven: no documented web SDK, and no first-party Flutter, React Native or Unity wrapper. Google’s own guidance also caps what you can combine. Contours are detected for only the most prominent face in an image, the Android guide advises against enabling contour detection and face tracking together, and the eyes-open and smiling classifiers work only on frontal faces within a Euler Y angle of -18 to 18 degrees.</p><h3>What Banuba covers, and what it leaves to you</h3><p>Scope first. Eight platforms from one integration, with iOS 13, Android 8.0, Windows 8.1, macOS 10.13 and any browser with WebGL 2.0 as the minimums. Our product page puts that at 97% of Apple and 80% of Android smartphones. Over 1,000 Asset Store filters plus Banuba Studio means the first release does not wait on bespoke effect authoring.</p><p>The boundaries are just as concrete. We do not publish a price list, so budget approval needs a sales conversation that DeepAR’s readers can skip. Our recommended three simultaneous faces on mobile sits below BytePlus’s documented ten. Our landmark count lives on the marketing page rather than in the technical document, which is the wrong way round. And a commercial license cannot be forked, which is the one thing MediaPipe gives you for nothing.</p><p>One worked example at scale: bemyfriends built the fandom SaaS platform b.stage on it, and <a href="https://www.banuba.com/blog/1-million-maus-in-2-years-for-a-saas-fandom-platform">b.stage passed 1 million monthly active users</a> in May 2024, two years after its April 2022 launch, with users in 224 countries.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*anYTAk-yw1aS_wc2Efjb6A.png" /><figcaption>B.STAGE by bemyfriends</figcaption></figure><h3>No provider here wins every criterion, so rank the criteria before you shortlist</h3><p>Shipping to iOS, Android, Web and Unity at once is a different problem from needing ten faces in one frame, from having no budget, and from having to own the source. Those four constraints point at four different names above.</p><p>Decide which one you cannot move, score the criteria against it, then ask the two or three survivors for the numbers their pages leave out. The question that separates them fastest is the one most of these pages do not answer: what frame rate, on which handset, with how many faces in shot.</p><h3>FAQ</h3><h4>Which AR SDK should a live streaming app choose?</h4><p>Score the criteria above against your own constraints, because the answer changes with them. Live streaming usually weights platform coverage, faces per frame and on-device frame rate highest, since group rooms put more than one face in shot and a dropped frame is visible to the audience.</p><h4>How much does Banuba Face AR SDK cost?</h4><p>There is no public price list. Banuba offers two models: a flat annual license fee with unlimited usage and a fixed cost, or an Active User Based model that costs less early and grows with the user base. A 14-day free trial is available on request, and the two models diverge sharply if the app scales, so raise the one you want in the first call.</p><h4>Can I use MediaPipe instead of a paid face tracking SDK?</h4><p>For landmark extraction, often yes: MediaPipe is Apache 2.0 and estimates 478 3D landmarks. What you take on is everything around it. Google labels the release an early preview, smoothing only applies at one face, there is no SLA or commercial support, and the project documents no makeup or try-on rendering, so effect work becomes your team’s build.</p><h4>Does Banuba Face AR SDK work offline?</h4><p>Yes. Every frame is processed on the user’s device, so there is no cloud round-trip and the SDK works offline as well as online. That removes the network latency budget and narrows the privacy review that a cloud-processing camera layer normally triggers.</p><h4>How many faces can these SDKs track at once?</h4><p>BytePlus documents the highest figure at 10 faces per frame. Banuba’s headline capability is five, and three is what we recommend on mobile, with our own specification showing the cost: with five faces in frame, the Android Low tier drops to 22 FPS and Android High to 27, while both iOS tiers hold 30. MediaPipe applies landmark smoothing only when the face count is set to 1, and Google ML Kit detects contours for the most prominent face only. DeepAR, Snap and Faceunity do not publish a maximum.</p><h4>How far can a face AR SDK track a face?</h4><p>It depends on the device, and most providers here publish nothing. Our own technical specification gives maximum face tracking distance as 170 cm on the Android Low tier, 180 cm on Android High and 230 cm on both iOS tiers. DeepAR publishes 150 to 180 cm. A 370 cm figure also appears in our specification, but it belongs to background separation in landscape orientation, which is a different capability from face tracking.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3a9fe7b0772b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AR Mirror Problems on the Store Floor: How to Tell Camera, Light, Kiosk and Tracking Apart]]></title>
            <link>https://banuba.medium.com/ar-mirror-store-floor-problems-435ff3440d95?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/435ff3440d95</guid>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[augmented-reality]]></category>
            <category><![CDATA[retail-technology]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 16:25:31 GMT</pubDate>
            <atom:updated>2026-10-28T16:25:31.383Z</atom:updated>
            <content:encoded><![CDATA[<h4>An AR smart mirror that behaves in the office and misbehaves on the store floor is failing in one of four layers that change between the two rooms: the camera, the light, the kiosk computer with its network, or face tracking at shopper distance.</h4><p>Each layer leaves its own symptom, and the symptom tells you which one to open first.</p><p>At a desk, the mirror runs on a laptop camera under even office light, with one developer sitting close to the lens. On a sales floor, a mounted webcam faces a shopper who stands further back, a window or a spotlight sits behind them, the screen runs from opening to closing, and a friend walks into the frame.</p><p>One symptom falls outside all four. A shade that looks wrong at the desk as well is a catalog problem, and moving the mirror will not fix it.</p><p>We build an AR mirror ourselves, so the product settings named below come from our own widget documentation, and nothing in this piece is a benchmark from a live store. The browser behavior and the camera behavior are the same whichever vendor you run.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*akrxU3i_zQZCwPD2m70ONQ.png" /></figure><h3>What does an AR mirror do with each camera frame?</h3><p>Every frame passes through the same chain. The camera captures it under the light the store provides, the browser hands it to the page, a tracking model finds the face, a renderer draws the products onto it, and the screen shows the result. A fault early in the chain can look like a fault further down.</p><pre>store light<br>  -&gt; camera          exposure, white balance, frame rate<br>  -&gt; browser         secure origin, camera permission, default camera<br>  -&gt; face tracking   face size in the frame, head angle, people in view<br>  -&gt; rendering       the kiosk&#39;s graphics chip and processor<br>  -&gt; screen</pre><pre>network: loads the try-on script, its settings and the catalog at startup</pre><p>The camera stage decides more than it seems to. A webcam adjusts exposure and white balance by itself, so the frames that reach the tracker change whenever the room changes, even when nothing in the code does.</p><p>The browser stage is a gate. MDN describes <a href="https://developer.mozilla.org/en-US/docs/Web/API/MediaDevices/getUserMedia">getUserMedia()</a> as “a powerful feature that can only be used in secure contexts”, which means a page served over plain HTTP never receives a camera at all.</p><p>Tracking and rendering do the heavy work, and the drawing happens on the kiosk’s own hardware. A setting that sharpens the picture spends frames to do it.</p><h3>Which symptom points to which layer?</h3><p>Match what staff report to the layer that produces it before you touch a single setting. A black or wrong view is the camera, a dark or shifting picture is the light, stutter is the kiosk, and slipping products are tracking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7pV_jxGKua3nzFEU39hMkg.png" /></figure><p>The same four lines in prose, for anyone reading this on a phone:</p><ul><li><strong>Camera<br></strong>A black screen, a permission prompt or a view from the wrong camera is settled in the browser’s camera settings, before any change to the light or the hardware. Start with which camera the browser handed to the page.</li><li><strong>Light<br></strong>A picture that is too dark, a bright halo around the shopper, or colors that drift during the day point at windows and spotlights, because the camera keeps re-adjusting to them. Start with what sits behind the shopper and how the light gate is set.</li><li><strong>Kiosk computer and network<br></strong>Choppy motion that survives good light belongs to the machine, and a slow first load or a late-day freeze belongs to the network and the build. Start with frames drawn per second, render resolution and the widget build in use.</li><li><strong>Face tracking at distance<br></strong>Products that hold at arm’s length and slip when a shopper steps back, turns, or is joined by a friend mean the face is too small in the frame for where shoppers actually stand. Start with how large the face is at that spot.</li></ul><h3>Is the AR mirror showing a black screen or the wrong camera?</h3><p>A black screen on a kiosk that worked at the office means the page has no camera feed. Check two settings first: whether the page has camera permission, and which camera the browser picks when the machine has more than one. Both live on the kiosk itself, and you can confirm both from the browser console in a minute.</p><p><a href="https://www.banuba.com/ar-mirror">Banuba’s AR mirror</a> is virtual try-on software that runs in the browser on in-store smart mirrors. Its camera call, useWebcam(), resolves with the active webcam stream, or with null if the user denied camera access, as the <a href="https://tintvto.com/docs/widget-api.html">widget documentation</a> describes. A null on a kiosk therefore points at permission before it points at the hardware.</p><p>Permission is granted per site. Chrome’s prompt offers “Allow while visiting the site”, “Allow this time” and “Never allow”, and its Site settings page lists the sites that are allowed or not allowed to use the camera. Pick the first option while you set the kiosk up, then confirm that the mirror’s address appears on the allowed list.</p><p>A managed browser can take this out of your hands. Chrome’s <a href="https://support.google.com/chrome/answer/2693767">help page</a> states that if you are using Chrome at work or school, your network administrator can set camera and microphone settings.</p><p>A kiosk with a camera built into the screen and a second one mounted on the frame gives the browser a choice. Chrome lets you set the default camera under Settings, Privacy and security, Site settings, Camera, and the drop-down there switches between devices.</p><p>To see which camera the widget actually received, run this on the kiosk with the console open:</p><pre>&lt;script type=&quot;module&quot; src=&quot;https://tintvto.com/widget.js&quot;&gt;&lt;/script&gt;<br>&lt;tint-vto merchant-id=&quot;YOUR_MERCHANT_ID&quot; skip-source&gt;&lt;/tint-vto&gt;</pre><pre>&lt;script type=&quot;module&quot;&gt;<br>  (async () =&gt; {<br>    const widget = document.querySelector(&quot;tint-vto&quot;);<br>    await widget.open();</pre><pre>// useWebcam() resolves with the active webcam stream,<br>    // or null if camera access was denied<br>    const stream = await widget.useWebcam();<br>    if (!stream) {<br>      console.warn(&quot;No camera: check this site&#39;s camera permission on the kiosk&quot;);<br>      return;<br>    }</pre><pre>// Standard MediaStream API. Read only: the widget owns this<br>    // track, so do not stop it.<br>    const [track] = stream.getVideoTracks();<br>    if (!track) {<br>      console.warn(&quot;Camera stream carries no video track&quot;);<br>      return;<br>    }</pre><pre>const { width, height, frameRate } = track.getSettings();<br>    console.log(&quot;Camera:&quot;, track.label, `${width}x${height}`, `${frameRate} fps`);<br>  })();<br>&lt;/script&gt;</pre><p>The skip-source attribute removes the “Live camera / Upload Photo” chooser, so the mirror goes straight to the camera when it opens. The track’s label is the browser’s own name for the camera source, which tells you whether the stream comes from the webcam on the mirror or the one inside the screen.</p><p>MDN describes <a href="https://developer.mozilla.org/en-US/docs/Web/API/MediaStreamTrack/getSettings">getSettings()</a> as returning “the current values of each of the constrainable properties” for the track. Its width, height and frame rate are the first numbers to compare when the picture stutters.</p><p>The same camera step looks different on a shopper’s own phone, where a photo can stand in for a denied camera, and our <a href="https://banuba-official.hashnode.dev/virtual-glasses-try-on-sdk-what-runs-between-the-camera-and-the-cart">walkthrough of a glasses try-on widget from camera to cart</a> covers that path. On a mirror, no shopper should ever see a camera permission prompt, so set the kiosk up so that it never appears.</p><h3>Is the picture too dark, or do the colors drift during the day?</h3><p>When a window, a doorway or a lit display sits behind the shopper, the camera exposes for that bright background and the face comes out dark. Colors drift for a related reason: daylight and store lamps have different color temperatures, and the camera keeps re-balancing between them as the sun moves and the lamps stay fixed.</p><p>The Linux kernel’s <a href="https://docs.kernel.org/userspace-api/media/v4l/ext-ctrls-camera.html">camera control reference</a> shows how far apart those sources sit. It puts the incandescent white balance preset at approximately 2500 to 3500 K, fluorescent at approximately 4000 to 5000 K, and daylight at approximately 5000 to 6500 K. A floor that mixes a warm spotlight with a window asks the camera to settle between two of those ranges at once.</p><p>Dim light can also cost frames. The same reference describes an exposure control that determines whether the device “may dynamically vary the frame rate”, and notes that the feature is disabled by default and the frame rate must remain constant. A webcam with it switched on can lower its frame rate to gather light in a dark corner, which staff will report as lag.</p><p>The Banuba virtual try-on widget has its own check on this layer. Its documentation describes the lighting-coefficient attribute as a numeric threshold for the lighting estimation gate, and states that values lower than the provided coefficient will block the user from continuing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*akS8qUvnmzThQ8Iof5ieYA.png" /></figure><p>The documentation gives no scale and no default, so adjust it in small steps with a colleague standing on the shoppers’ spot during the darkest hour of the store day.</p><p>The fix that lasts is physical. Turn the mirror so the window sits beside the camera, add a soft light that faces the shopper from the mirror side, and keep bright signage out of the camera’s view.</p><h3>Why is the AR mirror choppy, slow to start, or frozen by evening?</h3><p>Choppy motion that persists in good light belongs to the kiosk computer: the render resolution is too high for its graphics chip, or the camera delivers fewer frames than the page can draw. A slow first load belongs to the network. Check a late-day freeze against the widget build and the hours since the last reload.</p><p>Compare two frame rates. The camera’s rate comes from getSettings() in the snippet above. The page’s rate comes from Chrome DevTools, where the Rendering tab’s Frame rendering stats overlay shows a real-time estimate of frames per second, dropped frames in red, and GPU memory use.</p><p>If the camera holds its rate while the page drops frames, the machine is short of graphics power. If the camera itself reports a low rate, go back to the light.</p><p>The widget’s device-pixel-ratio attribute is the main render dial. The documentation sets its default at 1 and says that higher values improve sharpness but cost more GPU and CPU. A large mirror screen looks better at a higher value on the bench, so watch the overlay’s frame count each time you raise it on the kiosk itself.</p><p>Startup depends on the network. The widget loads as a script from Banuba’s CDN, as the <a href="https://tintvto.com/docs/integration.html">integration guide</a> describes, and the merchant ID then tells it to fetch its token, feature flags, translations and styling from the same place.</p><p>A store network that is slow or filtered shows up as a mirror that takes a long time to start. Open the DevTools Network panel on the kiosk at opening time and look for slow or blocked requests to that domain. The documented script URL carries no version number, and a team that needs a fixed build can ask Banuba for the customized widget archive the documentation mentions.</p><p>Browser age matters on a kiosk that never updates. The widget supports the last 5 versions of Chrome and Edge and the last 2 major versions of Safari, so a kiosk image frozen on an old browser drifts out of support without anyone touching it.</p><p>Long sessions have their own release history. The widget <a href="https://tintvto.com/docs/changelog.html">changelog</a> lists improved session reliability during longer usage in v1.5.0. If the store runs that customized archive, compare its version with the changelog before you blame the hardware, and read the live changelog first, because later releases can change the picture.</p><p>Teams that build the mirror as a native app use Banuba’s Face AR SDK instead, with sample projects for <a href="https://github.com/Banuba/banuba-sdk-ios-samples">iOS</a> and <a href="https://github.com/Banuba/banuba-sdk-android-samples">Android</a> on GitHub and setup guides in the <a href="https://docs.banuba.com/far-sdk/">Face AR SDK documentation</a>. The four layers stay the same, while camera selection and render resolution move from browser settings into your own code.</p><h3>Do products slip when shoppers step back, turn, or come in pairs?</h3><p>Face tracking needs enough of the face in the frame. A camera framed for arm’s length sees a much smaller face when the shopper steps back to see a whole look, and a small face gives the tracker less to hold. A turned head, a raised hand and a second shopper add to the load.</p><p>Banuba’s AR mirror page names three behaviors for this layer. Shoppers operate the mirror with gestures and never touch the screen. Tracking stays reliable when a face is partly covered, such as by a medical mask. Several shoppers can try products at the same mirror at once.</p><p>Density helps once the face is large enough, and does not rescue a face that is too small. On the makeup side, our <a href="https://www.banuba.com/tint-makeup-virtual-try-on">virtual makeup try-on page</a> states that face tracking runs on 3,308 vertices rather than a small set of landmarks.</p><p>We do not publish a working distance for the browser widget. Banuba’s face-tracking distance figures sit in the Face AR SDK technical specification, measured on Android and iOS phones, and the specification calls them reference values achieved under fixed lab conditions and asks teams to measure each feature in their own setting. The distance check on a sales floor is therefore yours to run.</p><p>Mark the spot where shoppers actually stand to see a full outfit or a pair of earrings, and have a colleague stand there while you watch the products. Move the camera, the mark or the screen until the products hold at that spot. Then repeat with two people in the frame and with a hand raised to the face.</p><p>Gestures depend on the same frame. A camera that struggles to see a face at the back mark will also struggle to see a hand there, so test the gestures from the mark as well.</p><h3>What should the mirror log before the next complaint?</h3><p>Staff describe symptoms in their own words, sometimes hours after they happen. A small log on the kiosk turns “the mirror was weird this afternoon” into a layer: which camera was active, at what size and frame rate, how many frames the page drew, which widget build was loaded, and what the light was doing at that hour.</p><p>Log these on every kiosk:</p><ul><li>The camera label, width, height and frame rate at each start, from the snippet above.</li><li>The page’s frame rate and dropped frames during the busiest hour, from the Frame rendering stats overlay.</li><li>The load time of the widget script and its settings at opening time, from the Network panel.</li><li>The widget build in use, checked against the changelog.</li><li>A time stamp for every staff complaint, so it can be lined up against daylight and the lamp schedule.</li><li>Products added to the cart from the mirror, through the widget’s addToCart event, so a quiet mirror can be told apart from a broken one.</li></ul><p>Crashes leave a trace on our side as well. The widget documentation says it ships with a built-in Sentry error tracker, so include the kiosk’s time stamps when you report a crash to us.</p><p>Some causes cannot be fixed on site, such as a window the store will not cover. The changelog lists a route around the mirror for those: a try-on link or QR code for a product or catalog, added in v1.7.1, so a shopper can carry on with the same try-on on their own phone.</p><h3>Fix the camera and the light first, then the kiosk and the floor plan</h3><p>The layers stack. A wrong camera or a backlit face makes the tracker look weak and the kiosk look slow, so new hardware bought before those two checks can leave the real fault in place.</p><p>The camera and the kiosk can be read from the console and DevTools without leaving the mirror. The light and the shoppers’ spot need someone on the floor at the darkest and busiest hours, so book that visit once the console checks are done.</p><h3>FAQ</h3><h4>Does Banuba’s AR mirror work with an external USB webcam?</h4><p>Banuba’s AR mirror uses whatever camera the browser hands it, and the browser treats a USB webcam as a camera like any built-in one. Set the USB webcam as Chrome’s default camera on the kiosk, then confirm it with the label check above. The <a href="https://tintvto.com/docs/">widget documentation</a> lists the supported browsers and the HTTPS rule but no camera models, so confirm your exact model with Banuba before a rollout.</p><h4>Is Banuba’s AR mirror real-time AR or a photo-based try-on?</h4><p>Banuba’s AR mirror is real-time AR: shoppers see the products on a live camera feed. The same widget can also take an uploaded image through its useImage() method, which the documentation suggests as the fallback when camera access is denied, but on a kiosk the live camera is the normal mode.</p><h4>Which products can shoppers try on at a Banuba AR mirror?</h4><p>Banuba’s AR mirror page lists makeup, jewelry, headwear, eyeglasses and sunglasses, watches, scarves, contact lenses and nail polish. For in-store hair color consultations, Banuba’s <a href="https://www.banuba.com/hair-ai">Hair AI SDK</a> adds hair color try-on, hairstyle try-on and shade analysis to the same mirror setup.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=435ff3440d95" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[6 steps of video editing app development, from concept to store launch]]></title>
            <link>https://banuba.medium.com/video-editing-app-development-a788c3978c09?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/a788c3978c09</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[video-editing]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 08:24:47 GMT</pubDate>
            <atom:updated>2026-10-27T08:24:47.253Z</atom:updated>
            <content:encoded><![CDATA[<h4>From audience to store listing, with a wage-bill calculation and the Apple and Google testing rules that set the launch date.</h4><p>Video editing app development is a sequence of six decisions: audience, revenue model, build route, tech stack, MVP scope and release. The build route sets most of the cost and the calendar, and three inputs price it: headcount, months and a median wage.</p><p>Banuba Video Editor SDK is a mobile video editing engine for iOS, Android, Flutter and React Native. It is ours, which places it among the routes in step 3.</p><p>None of the figures that follow is from our own measurements. They come from our documentation, our pricing guide, one customer case study and the third-party sources named beside them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4MjtbqZZR38dpUUdbovI0g.png" /></figure><h3>What do you need before video editing app development starts?</h3><p>Line up four things before anyone hires or signs:</p><ul><li>One named target user, specific enough that you could later recruit others like them.</li><li>The list of things that person must do with a clip on day one.</li><li>The platforms at launch: iOS, Android or both.</li><li>One person who owns the budget and the launch date.</li></ul><h3>Step 1: Who is the editor for, and what will they make with it?</h3><p>A fast and fun editor gets someone to a finished post in a few taps, with effects, music and captions carrying the result. An in-depth editor offers a creator a timeline, precise trims and control over each layer.</p><p>The choice at this stage is which of the two goes out in version one. At MVP you cannot be both, because each pulls the interface and the scope in opposite directions.</p><p>The trap is an editor for everyone: it inflates the backlog you will score in step 5 and pushes go-live further out.</p><h3>Step 2: How will a video editing app make money?</h3><p>Revenue options that suit an editor:</p><ul><li>A freemium subscription with paid tiers.</li><li>Watermark removal: unpaid files carry your mark, and subscribers get clean ones.</li><li>Ads, with a paid option to remove them.</li><li>A free companion to an offering you already sell.</li><li>Effect packs sold in the app.</li></ul><p>The model decides what must exist in release one. With watermark removal, the editor has to stamp unpaid output, and in Banuba Video Editor SDK a watermark is not added to exported video by default.</p><p>On Android, your engineers plug in your own watermark image through a custom provider, as our <a href="https://docs.banuba.com/ve-pe-sdk/docs/android/guide_export/">Android documentation</a> describes.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FH4GbHy-7KPY%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DH4GbHy-7KPY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FH4GbHy-7KPY%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/5defd4ee0997091fecc167a774e944d4/href">https://medium.com/media/5defd4ee0997091fecc167a774e944d4/href</a></iframe><p>Choosing how to earn after the backlog is frozen turns that stamp, or an ad slot, into a late change request.</p><h3>Step 3: Who builds it: an in-house team, an agency, freelancers or an SDK?</h3><p>This choice drives most of your budget, so begin with the wage bill of doing it yourself.</p><p>The headcount and timeline below are an illustrative planning assumption from the 2021 version of <a href="https://www.banuba.com/blog/how-to-develop-a-mobile-video-editor-app">our guide to building a mobile video editor app</a>, which estimated “at least 5 people and three months of work”. The wage is current and the team estimate is not, which is one more reason to swap in your own numbers. The arithmetic holds either way.</p><ol><li>Input 1, headcount: 5 developers.</li><li>Input 2, duration: 3 months.</li><li>Input 3, salary: $135,980, the median annual wage for software developers in the US in May 2025, according to the <a href="https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm">U.S. Bureau of Labor Statistics</a>.</li><li>Result: 5 x $135,980 x 3/12 = $169,975 in wages alone, before hiring, taxes, benefits, equipment and QA.</li></ol><p>Path Who writes the editor What drives the cost What you own afterwards Main risk In-house native team Your iOS and Android hires Double headcount, for creation and upkeep All of it, with the know-how Hiring delays, double maintenance In-house cross-platform team A single group on a shared codebase One payroll, plus native help All of it Recording and effects still require native experts Outsourced agency An outside firm Its rate for the project length The source, if the agreement transfers it Communication gaps, IP terms, dependence on one supplier Freelance team Contractors you coordinate Hourly rates and your management time The same, if each deal transfers it Hard to manage and to replace Mixed team Your staff plus outside specialists Salaries and specialist fees Most of it Split responsibility for defects SDK inside your own app The SDK company; your side wraps the app around it A yearly fee plus setup Your app and design; the engine stays licensed Supplier terms, no web editor</p><p>The cells are our own summary of the trade-offs and hold no measured figures. Each row again, in text:</p><h4><strong>In-house native team</strong></h4><p>Your own iOS and Android hires write it, you fund double headcount for creation and upkeep, you own all of it along with the know-how, and the risk is hiring delays plus maintaining two codebases. Makes sense when the editor is the core of the business and you can fund both hires for years.</p><h4><strong>In-house cross-platform team</strong></h4><p>A single group works on a shared codebase, you carry one payroll plus native help, you own all of it, and the risk is that recording and effects still need native experts. Suits a company with Flutter or React Native skills in place that can also buy native help.</p><h4><strong>Outsourced agency</strong></h4><p>An outside firm writes it at its rate for the length of the project, you own the source if the agreement transfers it, and the risks are communication gaps, IP terms and dependence on one supplier. The right call when you have no engineers of your own and the agreement transfers ownership to you.</p><h4><strong>Freelance team</strong></h4><p>Contractors you coordinate, billed hourly plus your management time, ownership transferring only if each individual deal says so, and the risk is that they are hard to manage and to replace. Works for a short, well-defined scope that someone on your side can supervise closely.</p><h4><strong>Mixed team</strong></h4><p>Your staff plus outside specialists, paid in salaries and specialist fees, with you owning most of it and carrying split responsibility for defects. Pays off when your people own the roadmap and outside specialists cover the media pieces.</p><h4><strong>Licensed SDK inside your own app</strong></h4><p>The SDK company writes the editor and your side wraps the app around it, the cost is a yearly fee plus setup, you own your app and design while the engine stays licensed, and the risks are supplier terms and the absence of a web editor. Fits when video creation is one part of a bigger app and the calendar matters more than owning the engine.</p><p>For that last path, <a href="https://www.banuba.com/video-editor-sdk">Banuba Video Editor SDK</a> covers iOS, Android, Flutter and React Native, and every platform you ship on sits inside one license.</p><p>Our pricing guide puts integration at about 8 minutes using our agent skills for AI coding assistants, a few hours by hand, or 1 to 2 days for a fully customized build. The first of those three is the one with a condition attached, so read it as the floor rather than the expectation.</p><p>If you are weighing vendors against each other, start from <a href="https://medium.com/@banuba/7-best-mobile-video-editing-sdks-feature-by-feature-comparison-1b3519c8592">the seven mobile editing SDKs we lined up side by side</a>.</p><p>Every choice gives something up. An in-house group wins on ownership and control of the code, and recruiting is what slows it down.</p><p>An agency or freelancers win when you have no team and need people fast. The downside is communication overhead, IP terms to negotiate, reliance on outsiders, and contractors who are tough to manage and replace.</p><p>Our SDK has limits too. Web is not supported for the editor, it is licensed per app on a yearly contract, and extras such as custom features are quoted separately.</p><p>So, in short:</p><ul><li>If the editor is the thing customers buy, assign a native crew of your own to it, because you then hold every roadmap call.</li><li>If a whole crew has to exist by next quarter, go with an agency, because it skips the hiring round.</li><li>If your engineers already know Flutter or React Native, pair them with a ready-made editor, because shared code stops at the recording layer.</li><li>If video is a single tab in a social or commerce app, take an SDK, because your engineers stay on the features that make the app yours.</li></ul><h3>Step 4: Native or cross-platform, and where does a video editor SDK fit?</h3><p>Native means writing in Swift for iOS and Kotlin for Android. It gives the best camera access and the familiar look of each system, and it doubles the codebases you fund.</p><p>Cross-platform frameworks such as Flutter and React Native share one codebase between iOS and Android. Camera and editing features still need native work underneath, or a premade SDK that already contains it.</p><p>Frameworks that wrap a web page in an app shell, such as Ionic or Cordova, are a poor fit for a camera-first editor.</p><p>The trap here is choosing cross-platform to save on payroll, then rebuilding the camera layer natively anyway and staffing it twice.</p><p>Banuba Video Editor SDK provides official Flutter and React Native plugins for hybrid apps. The minimums are iOS 15.0 or newer, and Android 8.0 or newer with the Camera 2 API and Kotlin 2.1 or Java 17.</p><p>Size is worth reading per platform rather than carrying one number across both. Per Banuba’s Android requirements, the core editor adds 25 MB to your app’s download size and the full bundle 45 MB. The iOS figures are 14 MB for the core editor and the same 45 MB for the full bundle.</p><p>Video processing happens on the phone itself, and the algorithms work offline.</p><p>If iOS comes first for you, we mapped <a href="https://dev.to/team_banuba/6-routes-for-adding-a-video-editing-sdk-to-an-ios-app-and-where-each-breaks-1053">the six routes for adding a video editing SDK to an iOS app</a> and where each one breaks.</p><h3>Step 5: Which features go into a video editor MVP?</h3><p>The Value versus Complexity model from <a href="https://www.productplan.com/learn/strategies-prioritize-product-features">ProductPlan</a> scores every candidate on business value and on the effort to implement it. The high-value, low-effort items ship now, and the rest wait for a later release.</p><p>Weat, a US-based food-centric, video-native social commerce platform, is one example. Its MVP editor covered trimming, recording including a hands-free mode, audio editing, captions, picture and GIF overlays, AR face filters, LUT color filters, drafts and picture-in-picture.</p><p>Nine items, and worth noticing what is not among them: no multi-track timeline, no keyframes, no desktop companion.</p><p>In <a href="https://www.banuba.com/blog/video-editor-sdk-weat-case-study">Banuba’s published Weat case study</a>, Weat’s CTO says the SDK cut their development effort in half, and Weat launched on iOS in November 2021. That half is one customer’s own statement in a story we published, and it carries no weight as an industry average.</p><p>What goes wrong here is an item scored as valuable while nobody has sized the job behind it. Ask whoever will own each item for an estimate before the backlog is final.</p><h3>Step 6: How do you test and launch a video editing app in the stores?</h3><p>Acceptance testing has four ordered parts:</p><ol><li>Write success criteria from your user stories.</li><li>Turn them into test cases such as “add a music track and export”.</li><li>Recruit testers from the target user group and keep your own staff out of the panel.</li><li>Check every result against the criteria from part 1.</li></ol><p>On iOS, <a href="https://developer.apple.com/testflight/">Apple’s TestFlight</a> lets you invite up to 10,000 external testers, and your first build has to be approved by App Review for TestFlight.</p><p>Android is where the calendar actually gets decided. Per <a href="https://support.google.com/googleplay/android-developer/answer/14151465">Google’s Play Console rules</a>, a personal developer account created after 13 November 2023 must run a closed test with at least 12 testers opted in for the preceding 14 consecutive days before it can apply for production access. A tester who opts out and back in restarts the count. Google states that review usually takes seven days or less.</p><p>For a new personal Android account, the minimum calendar looks like this:</p><ol><li>Closed test: 14 days, per Google’s rule.</li><li>Review: up to 7 more, per Google’s usual window.</li><li>Total: 14 + 7 = 21 days, about 3 weeks between a test-ready Android version and production access.</li></ol><p>A launch date fixed before that closed-test clock started is the failure at this stage, so plan that window into the date.</p><p>Time an export of a one-minute clip on the oldest phone your audience carries, using our <a href="https://www.banuba.com/blog/ve-sdk-performance-benchmark-compared">video editor SDK performance benchmark</a> as a reference point.</p><p>The App Store listing has hard limits, per <a href="https://developer.apple.com/app-store/product-page/">Apple’s product page guidance</a>: an app name of up to 30 characters, a subtitle of up to 30, a keywords field of 100 characters, and up to three app previews of up to 30 seconds each.</p><p>For a video editing app the preview is the product demo, so it should show the editor in use.</p><h3>Settle who writes the editor, then start the test clock early</h3><p>Of the six steps, step 3 sets most of the cost, because it decides whom you pay and for how long.</p><p>In elapsed time, the illustrative in-house line from step 3 is 3 months of a 5-person team. On the SDK path, Banuba’s pricing guide puts integration at a few hours by hand to 1 to 2 days fully customized, and says most teams take two to three weeks from first contact to a signed contract, with a 14-day trial running in parallel.</p><p>Either way, the Play closed test adds about 3 weeks, and that clock only starts once there is something to test.</p><h3>FAQ</h3><h4>How much does it cost to develop a video editing app with Banuba Video Editor SDK?</h4><p>There is no public price list. Banuba Video Editor SDK pricing is MAU-based and depends on two numbers: the monthly active users your app carries and the platforms you ship on. Banuba licenses it per app on a yearly contract, and the amount stays fixed for that year, with multiple apps able to sit in one contract.</p><p>Banuba charges no per-developer seat fee, quotes add-ons such as integration assistance separately, and offers a 14-day free trial covering the full feature set. The <a href="https://www.banuba.com/blog/banuba-video-editor-sdk-pricing-guide">Banuba Video Editor SDK pricing guide</a> walks through the details.</p><h4>How hard is it to make a video editing app with Banuba Video Editor SDK?</h4><p>Banuba puts the setup at about 8 minutes using its agent skills, a few hours by hand, and 1 to 2 days for a fully customized build.</p><p>The agent skills for the editor install into Claude Code, Codex and Qwen Code in one command, as the <a href="https://docs.banuba.com/ve-pe-sdk/docs/vibe-coding/">Banuba agent skills documentation</a> explains.</p><h4>Can you build a video editor app in Android Studio with Banuba Video Editor SDK?</h4><p>Yes. Banuba Video Editor SDK supports Android 8.0 or newer with the Camera 2 API, and Kotlin 2.1 or Java 17.</p><p>Banuba publishes an <a href="https://github.com/Banuba/ve-sdk-android-integration-sample">Android sample project on GitHub</a> as a reference.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a788c3978c09" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[iOS video editor SDK vs AVFoundation in Swift: 8 steps to build, 6 to integrate]]></title>
            <link>https://banuba.medium.com/ios-video-editor-sdk-swift-b444565f1de1?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b444565f1de1</guid>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[ios-development]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[swift]]></category>
            <category><![CDATA[video-editing]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 08:05:13 GMT</pubDate>
            <atom:updated>2026-10-27T08:05:13.357Z</atom:updated>
            <content:encoded><![CDATA[<h4>An iOS app gets a video editor one of two ways: AVFoundation code you write yourself, or an iOS video editor SDK you install as a dependency.</h4><p>The honest version of the comparison is not minutes against months. Installing a dependency takes minutes either way. What differs is everything after the install: the timeline, the effects pipeline, the preview and the export screens are either on your roadmap or they are not.</p><p>We are Banuba, and we build the SDK used in the second half, so read that half knowing who wrote it. The first half uses Apple frameworks only.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HeUKuCnbwZh4sOo3q3gKKA.png" /></figure><h3>What you need before you start</h3><p>Both routes need the same base setup:</p><ul><li>A Mac with Xcode and Swift installed.</li><li>Two short test clips, one for trimming and one for the cross-fade.</li><li>A music file for the audio step.</li></ul><p>The SDK route adds three items:</p><ul><li>Xcode 26.0 or newer and Swift 5.9 or newer, as the Banuba docs list them.</li><li>CocoaPods 1.12.1 or newer, or Swift Package Manager.</li><li>A trial license token from the Video Editor SDK trial form on the Banuba site.</li></ul><h3>How do you build a video editor on AVFoundation in Swift?</h3><p>A Swift video editor rests on five native pieces: AVFoundation, AVMutableComposition, AVVideoComposition with Core Image, AVPlayer, and AVAssetExportSession. <a href="https://developer.apple.com/documentation/avfoundation">Apple’s AVFoundation documentation</a> describes the framework that ties them together.</p><p>One note before the code. The snippets below use the synchronous AVAsset accessors, because they read more clearly in a tutorial. Apple deprecated them in iOS 16 in favour of the async load(_:) family, so Xcode will warn on them.</p><p>Before any of this goes near production, swap asset.tracks(withMediaType: .video) for try await asset.loadTracks(withMediaType: .video) and asset.duration for try await asset.load(.duration). The force unwraps are here for brevity too, and they are the first thing to replace.</p><h3>Step 1: import the frameworks</h3><p>AVKit supplies the player screen. Core Image supplies the filters.</p><pre>import AVFoundation<br>import AVKit<br>import CoreImage</pre><h3>Step 2: load a clip and trim it</h3><p>AVURLAsset is the file and AVAssetTrack is a track inside it. AVMutableComposition is the editable timeline, and addMutableTrack creates a track you control.</p><pre>let asset = AVURLAsset(url: videoURL)<br>let comp = AVMutableComposition()<br>guard let vSrc = asset.tracks(withMediaType: .video).first else { return }<br>let vDst = comp.addMutableTrack(withMediaType: .video,<br>                                preferredTrackID: kCMPersistentTrackID_Invalid)!<br>let sixSec = CMTime(seconds: 6, preferredTimescale: 600)<br>try vDst.insertTimeRange(CMTimeRange(start: .zero, duration: sixSec),<br>                         of: vSrc, at: .zero)</pre><h3>Step 3: merge a second clip with a cross-fade</h3><p><a href="https://developer.apple.com/documentation/avfoundation/avvideocomposition">AVVideoComposition</a> holds time-based instructions that tell AVFoundation which pixels to render at each moment.</p><p>Place clip B so it overlaps the end of clip A by one second. Opacity ramps down on clip A and up on clip B, and both layer instructions sit inside one instruction covering the overlap.</p><p>Here is the part that catches people. The instructions have to cover the <strong>entire</strong> timeline, not only the dissolve. An instruction list containing just the overlap gives you a video that is blank everywhere else, and the symptom looks like a broken export rather than a missing instruction.</p><pre>let asset2 = AVURLAsset(url: videoURL2)<br>guard let vSrc2 = asset2.tracks(withMediaType: .video).first else { return }<br>let vDst2 = comp.addMutableTrack(withMediaType: .video,<br>                                 preferredTrackID: kCMPersistentTrackID_Invalid)!</pre><pre>let overlapDur = CMTime(seconds: 1, preferredTimescale: 600)<br>let tEndFirst = vDst.timeRange.end<br>let tStart2 = CMTimeSubtract(tEndFirst, overlapDur)<br>try vDst2.insertTimeRange(CMTimeRange(start: .zero, duration: asset2.duration),<br>                          of: vSrc2, at: tStart2)</pre><pre>// 1. Clip A alone, from the start to the dissolve.<br>let before = AVMutableVideoCompositionInstruction()<br>before.timeRange = CMTimeRange(start: .zero, duration: tStart2)<br>before.layerInstructions = [AVMutableVideoCompositionLayerInstruction(assetTrack: vDst)]</pre><pre>// 2. The dissolve itself.<br>let overlap = AVMutableVideoCompositionInstruction()<br>overlap.timeRange = CMTimeRange(start: tStart2, duration: overlapDur)<br>let a = AVMutableVideoCompositionLayerInstruction(assetTrack: vDst)<br>let b = AVMutableVideoCompositionLayerInstruction(assetTrack: vDst2)<br>a.setOpacityRamp(fromStartOpacity: 1, toEndOpacity: 0, timeRange: overlap.timeRange)<br>b.setOpacityRamp(fromStartOpacity: 0, toEndOpacity: 1, timeRange: overlap.timeRange)<br>overlap.layerInstructions = [a, b]</pre><pre>// 3. Clip B alone, from the end of the dissolve to the end of the timeline.<br>let afterStart = CMTimeAdd(tStart2, overlapDur)<br>let after = AVMutableVideoCompositionInstruction()<br>after.timeRange = CMTimeRange(start: afterStart,<br>                              duration: CMTimeSubtract(comp.duration, afterStart))<br>after.layerInstructions = [AVMutableVideoCompositionLayerInstruction(assetTrack: vDst2)]</pre><pre>let vcomp = AVMutableVideoComposition()<br>vcomp.instructions = [before, overlap, after]<br>vcomp.renderSize = vSrc.naturalSize<br>vcomp.frameDuration = CMTime(value: 1, timescale: 30)</pre><p>One more thing about that renderSize. naturalSize ignores the track&#39;s preferredTransform, so a clip recorded in portrait renders sideways. Read the transform and swap width and height when the rotation calls for it, or every phone-shot video comes out of your editor on its side.</p><h3>Step 4: apply a real-time filter</h3><p><a href="https://developer.apple.com/documentation/coreimage">Core Image</a> offers GPU-accelerated CIFilter objects, such as CIColorControls, CISepiaTone and CIVignette. They chain without custom Metal shaders.</p><p>Here is the first real decision point. The cross-fade composition from Step 3 and this CI-filtered composition are separate objects, and a player item takes only one of them. Combining a dissolve with a Core Image filter means rendering the filter inside the composition rather than alongside it.</p><pre>let ciContext = CIContext()<br>let noir = CIFilter(name: &quot;CIPhotoEffectNoir&quot;)!<br>let filteredVC = AVVideoComposition(asset: comp) { request in<br>    noir.setValue(request.sourceImage.clampedToExtent(), forKey: kCIInputImageKey)<br>    let out = (noir.outputImage ?? request.sourceImage)<br>        .cropped(to: request.sourceImage.extent)<br>    request.finish(with: out, context: ciContext)<br>}</pre><h3>Step 5: add music with a fade-in</h3><p>The music track joins the same composition, and its volume ramps up over two seconds. For live voiceover, EQ, reverb or pitch, plan on AVAudioEngine instead.</p><p>Declare the mix outside the optional binding, or Step 6 cannot see it.</p><pre>let audioMix = AVMutableAudioMix()<br>let music = AVURLAsset(url: musicURL)</pre><pre>if let mTrack = music.tracks(withMediaType: .audio).first {<br>    let aDst = comp.addMutableTrack(withMediaType: .audio,<br>                                    preferredTrackID: kCMPersistentTrackID_Invalid)!<br>    try aDst.insertTimeRange(CMTimeRange(start: .zero, duration: comp.duration),<br>                             of: mTrack, at: .zero)<br>    let params = AVMutableAudioMixInputParameters(track: aDst)<br>    params.setVolumeRamp(fromStartVolume: 0, toEndVolume: 1,<br>                         timeRange: CMTimeRange(start: .zero,<br>                                                duration: CMTime(seconds: 2,<br>                                                                 preferredTimescale: 600)))<br>    audioMix.inputParameters = [params]<br>}</pre><h3>Step 6: preview the edit</h3><p>AVPlayer is the bare playback engine with no UI. AVPlayerViewController adds Apple’s standard transport UI, Picture-in-Picture, AirPlay and accessibility.</p><p>Many teams prototype with it, then move to a custom AVPlayer UI as requirements grow.</p><pre>let item = AVPlayerItem(asset: comp)<br>item.videoComposition = filteredVC   // or vcomp<br>item.audioMix = audioMix<br>let player = AVPlayer(playerItem: item)<br>let pvc = AVPlayerViewController()<br>pvc.player = player<br>present(pvc, animated: true) { player.play() }</pre><h3>Step 7: bind a slider timeline</h3><p>A UISlider seeks by converting its value to CMTime, and a periodic time observer keeps it in sync.</p><p>Two things go wrong here in practice. The observer has to guard against a non-finite duration before the item is readyToPlay, and it keeps firing after the screen closes unless you call player.removeTimeObserver(token) on exit.</p><pre>let interval = CMTime(seconds: 0.25, preferredTimescale: 600)<br>let token = player.addPeriodicTimeObserver(forInterval: interval, queue: .main) { [weak player] t in<br>    guard let item = player?.currentItem else { return }<br>    let dur = item.duration.seconds<br>    guard dur.isFinite, dur &gt; 0 else { return }<br>    slider.value = Float(t.seconds / dur)<br>}</pre><pre>@IBAction func sliderChanged(_ s: UISlider) {<br>    guard let item = player.currentItem else { return }<br>    let dur = item.duration.seconds<br>    guard dur.isFinite, dur &gt; 0 else { return }<br>    let seconds = Double(s.value) * dur<br>    player.seek(to: CMTime(seconds: seconds, preferredTimescale: 600),<br>                toleranceBefore: .positiveInfinity,<br>                toleranceAfter: .positiveInfinity)<br>}</pre><p>A richer timeline is a horizontal UICollectionView with AVAssetImageGenerator thumbnails, a fixed playhead and draggable trim handles.</p><p>Most teams go hybrid at this point: SwiftUI for app chrome and flows, UIKit for the timeline. <a href="https://sendbird.com/developer/tutorials/swiftui-vs-uikit">Sendbird’s SwiftUI vs UIKit breakdown</a> reaches the same split.</p><h3>Step 8: export the result</h3><p>Create a session with a preset, attach the same composition you previewed, and write the file. An existing file at the output URL makes the export fail or write a wrong file, which is why the code removes it first.</p><p>Export is also where the API changed. export(to:as:) arrived in iOS 18, and exportAsynchronously is the older call that still carries deployment targets below it. Branch on availability and both paths compile.</p><pre>let outURL = FileManager.default.temporaryDirectory<br>    .appendingPathComponent(&quot;final.mov&quot;)<br>try? FileManager.default.removeItem(at: outURL)</pre><pre>guard let session = AVAssetExportSession(<br>    asset: comp,<br>    presetName: AVAssetExportPresetHighestQuality<br>) else { return }</pre><pre>session.videoComposition = filteredVC   // or vcomp<br>session.audioMix = audioMix</pre><pre>if #available(iOS 18.0, *) {<br>    try await session.export(to: outURL, as: .mov)<br>} else {<br>    session.outputURL = outURL<br>    session.outputFileType = .mov<br>    await withCheckedContinuation { continuation in<br>        session.exportAsynchronously { continuation.resume() }<br>    }<br>    switch session.status {<br>    case .completed: break<br>    case .cancelled: throw CancellationError()<br>    default: throw session.error ?? ExportError.failed<br>    }<br>}</pre><h3>When does the native build stop being cheap?</h3><p>Core Image and AVFoundation are enough for a basic editor. The switch to <a href="https://developer.apple.com/documentation/metal">Metal</a> comes at the preview: move to it when the preview cannot hold 60 FPS, or at least 30 FPS, while users edit or overlay heavier effects.</p><p>Two typical triggers are live skin smoothing running at the same time as a film-style color LUT, and a sticker locked on a moving face.</p><p>Measure in Xcode Instruments with Metal System Trace and watch FPS and GPU time per frame. The budget is arithmetic you can redo: 1000 ms divided by 60 gives about 16.7 ms per frame at 60 FPS, and 1000 divided by 30 gives about 33.3 ms at 30 FPS.</p><p>Above the budget, downscale the preview while editing, then restore full quality on play.</p><p>The time cost follows the same curve. In our experience a junior iOS developer prototypes the stack in days and polishes it in weeks, and a polished editor with custom filters, effects and AR runs to months.</p><p>Cleveroad, a development agency that sells app builds, estimates that a video editing app starts from 670 engineering hours. The Cleveroad estimate adds a budget of $33,500 and counting, so both figures come from an interested party.</p><p>As an illustrative assumption, here is that effort as elapsed time:</p><ol><li>Effort input: 670 engineering hours, from Cleveroad.</li><li>Capacity input: one engineer at 40 hours a week.</li><li>Division: 670 / 40 = about 17 weeks.</li><li>Calendar view: 17 weeks is roughly four months.</li></ol><h3>Native component vs SDK module, stage by stage</h3><p>Each stage pairs a native component with an SDK module.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q7qobzY1ZsaHyvkBjhTrtw.png" /></figure><p>Native media loading is short AVAsset code, while the SDK also ships a recording screen.</p><p>On the native route, a thumbnail timeline with draggable trim handles is UI code your team writes.</p><p>Core Image covers basic native filters, while the SDK adds LUTs, masks, GIFs and a music catalog.</p><p>AVPlayerViewController gives a fast native preview, and a custom transport UI becomes extra work later.</p><p>Native export means configuring AVAssetExportSession yourself, while the SDK exports several files from one configuration.</p><p>If you want to see alternatives to Banuba before choosing, read <a href="https://medium.com/@banuba/7-best-mobile-video-editing-sdks-feature-by-feature-comparison-1b3519c8592">a feature-by-feature comparison of seven mobile video editing SDKs</a>.</p><h3>How do you integrate the Banuba iOS video editor SDK?</h3><p>The <a href="https://www.banuba.com/video-editor-sdk">Banuba Video Editor SDK</a> is a mobile video editing engine for iOS, Android, Flutter and React Native. It adds a TikTok-style editor, with native Swift integration on iOS and official plugins for Flutter and React Native.</p><p>As the <a href="https://docs.banuba.com/ve-pe-sdk/docs/ios/requirements/">Banuba docs requirements page</a> lists them, the SDK needs Swift 5.9, Xcode 26.0 and iOS 15.0 or newer. It supports arm64 and arm64e, with limited x86–64 support. Requirement pages change, and the live page decides.</p><p>The same page lists size. The core editor adds 14 MB to the app download size, the full bundle is 45 MB, Video Editor with Photo Editor is 80 MB, and Video Editor with face AR features runs 45 to 140 MB depending on the features enabled.</p><h3>Step 1: install the pods</h3><p>The <a href="https://docs.banuba.com/ve-pe-sdk/docs/ios/integration-ve/">Banuba docs integration guide</a> requires CocoaPods 1.12.1 or newer. Install it with brew install cocoapods and run pod init. The full dependency list sits in the Podfile of the <a href="https://github.com/Banuba/ve-sdk-ios-integration-sample">iOS integration sample on GitHub</a>.</p><p>The last two pods are optional. Decide on them now, since enabled features drive the final size.</p><pre>source &#39;https://cdn.cocoapods.org/&#39;<br>source &#39;https://github.com/Banuba/specs.git&#39;<br>source &#39;https://github.com/sdk-banuba/banuba-sdk-podspecs.git&#39;</pre><pre>banuba_sdk_version = &#39;1.52.1&#39;</pre><pre>pod &#39;BanubaVideoEditorSDK&#39;, banuba_sdk_version<br>pod &#39;BanubaSDKSimple&#39;, banuba_sdk_version<br>pod &#39;BanubaSDK&#39;, banuba_sdk_version<br>pod &#39;BanubaARCloudSDK&#39;, banuba_sdk_version      # optional<br>pod &#39;BanubaAudioBrowserSDK&#39;, banuba_sdk_version # optional</pre><p>Run pod install --repo-update, then open the generated .xcworkspace in Xcode and run.</p><p>Swift Package Manager is the alternative: the Banuba docs point to the sample’s spm branch, and that guide applies to VE SDK v1.31.3 and newer.</p><h3>Step 2: add permissions and strings</h3><p>Add four usage descriptions to Info.plist, for the media library, camera, microphone and photo library.</p><pre>&lt;key&gt;NSAppleMusicUsageDescription&lt;/key&gt;<br>&lt;string&gt;This app requires access to the media library&lt;/string&gt;<br>&lt;key&gt;NSCameraUsageDescription&lt;/key&gt;<br>&lt;string&gt;This app requires access to the camera.&lt;/string&gt;<br>&lt;key&gt;NSMicrophoneUsageDescription&lt;/key&gt;<br>&lt;string&gt;This app requires access to the microphone.&lt;/string&gt;<br>&lt;key&gt;NSPhotoLibraryUsageDescription&lt;/key&gt;<br>&lt;string&gt;This app requires access to the photo library.&lt;/string&gt;</pre><p>Then add the English Localizable.strings file from the sample to your project.</p><h3>Step 3: create the module and its configuration</h3><p>A VideoEditorModule class creates the VideoEditorConfig that the editor starts with.</p><pre>class VideoEditorModule {<br>    func createConfiguration() -&gt; VideoEditorConfig {<br>        var config = VideoEditorConfig()<br>        // ...<br>        return config<br>    }<br>}</pre><h3>Step 4: pass the license token and check its state</h3><p>The editor takes your license token and the configuration.</p><pre>let videoEditorSDK = BanubaVideoEditor(<br>    token: AppDelegate.licenseToken,<br>    configuration: config,<br>    externalViewControllerFactory: viewControllerFactory<br>)</pre><p>A bad token shows up right here: videoEditorSDK is nil when the token is empty or truncated. Check the license state before starting.</p><pre>videoEditorSDK?.getLicenseState(completion: { [weak self] isValid in<br>    if isValid {<br>        print(&quot;License is active&quot;)<br>    } else {<br>        print(&quot;License is either revoked or expired&quot;)<br>    }<br>    completion(isValid)<br>})</pre><p>With a revoked or expired license, the editor shows a “Video content unavailable” screen.</p><h3>Step 5: launch from the camera</h3><p>The camera entry point opens the recording screen first.</p><pre>let cameraLaunchConfig = VideoEditorLaunchConfig(<br>    entryPoint: .camera,<br>    hostController: self,<br>    musicTrack: nil,<br>    animated: true<br>)</pre><pre>videoEditorModule.presentVideoEditor(with: cameraLaunchConfig)</pre><h3>Step 6: export</h3><p>Create an ExportConfiguration object and add one or more ExportVideoConfiguration objects, each standing for a video or audio file. Then call BanubaVideoEditor.export() with it. The sample&#39;s ViewController.swift holds the full implementation.</p><h3>What you get and what you give up</h3><p>The recording screen covers a beauty filter, face and color filters, background music, recording speed, hands-free mode, zoom, flashlight, timer and mute. Users can capture several videos in one session, and progress is saved automatically.</p><p>The editing screen covers trim, music, voice recording and voice change, LUT color filters, text, GIFs, video effects, slow-mo, fast-forward and masks.</p><p>The product page also lists Picture-in-Picture, Templates, AI Clipping and AI Captions, with English, Mandarin, Spanish and Portuguese captions by default. A watermark on export is configurable, so you can place your own or leave it off.</p><p>The Audio Browser is an optional module. Banuba ships over 35 GB of royalty-free tracks in it by default, with a native Soundstripe integration, and it accepts other music APIs or uploads from the device.</p><p>Face AR is a second optional module, with try-on, morphing, character filters, a video background remover, 2D and 3D stickers and trigger effects.</p><p>Customization covers icons, colors, text styles, your logos, localized strings, export resolutions and durations, and recording duration behavior.</p><p>The scope ends at the screen flow. The UI version fixes the screen order and layout, because its modules have interconnected dependencies. The API version gives up some premade pieces, such as progress bars, and allows deeper customization.</p><p>AVFoundation keeps one clear advantage: full control of the pipeline, including HDR, Dolby Vision and ProRes. That control matters for a professional editing tool that ships only on iOS.</p><p>The SDK also ships agent skills for Claude Code, Codex and Qwen Code, each installed with a single command.</p><h3>What shipping with the SDK looked like for two apps</h3><p>Weat is a US social app described as TikTok for foodies. The <a href="https://www.banuba.com/blog/video-editor-sdk-weat-case-study">Weat case study</a> reports a November 2021 iOS launch after saving 50% of product development time.</p><p>Chingari is an Indian video sharing platform that integrated the editor to mimic TikTok features. The <a href="https://www.banuba.com/blog/video-editor-with-animation-effects-for-chingari-app-case-study">Chingari case study</a> reports 550,000 downloads in 10 days and more than 30 million in 3 months.</p><p>Both figures come from the customers rather than from a test we ran.</p><h3>Build on AVFoundation when editing is the product, integrate when it is a feature</h3><p>If the editor itself is what users come for, the native route earns its cost, and the control over HDR and ProRes stays with you.</p><p>If editing is one feature inside a social or UGC app, integration takes the timeline, effects and export screens off your roadmap. Your team keeps its months for the feed.</p><p>Either way, prototype the eight native steps first. A working trim, dissolve and export sizes the remaining work better than any estimate.</p><h3>FAQ</h3><h4>What happens when the Banuba Video Editor SDK license token is wrong, revoked or expired?</h4><p>A wrong token makes the Banuba Video Editor SDK initializer return nil, and empty or truncated tokens count as wrong.</p><p>A revoked or expired license shows a “Video content unavailable” screen, so call getLicenseState before launch. Token questions go to <a href="https://www.banuba.com/support">Banuba support</a>.</p><h4>Can an iOS editor built on the Banuba Video Editor SDK carry our own branding?</h4><p>Yes, the Banuba Video Editor SDK is offered as a white label editor with your logos and styles. You can change icons, colors, text styles and localized strings.</p><p>The UI version fixes the screen order, while the API version allows deeper customization. Banuba describes both versions on its <a href="https://www.banuba.com/white-label-video-editor">white label video editor page</a>.</p><h4>Does the Banuba Video Editor SDK ship agent skills for AI coding assistants?</h4><p>Yes, the Banuba Video Editor SDK ships agent skills for Claude Code, Codex and Qwen Code. Each one installs with a single command. Banuba keeps the setup on its <a href="https://docs.banuba.com/ve-pe-sdk/docs/vibe-coding">agent skills docs page</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b444565f1de1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Augmented Reality eCommerce: The Line Item Vendors Don’t Quote]]></title>
            <link>https://banuba.medium.com/augmented-reality-ecommerce-hidden-cost-ade32f4928f1?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ade32f4928f1</guid>
            <category><![CDATA[3d-modeling]]></category>
            <category><![CDATA[retail-technology]]></category>
            <category><![CDATA[augmented-reality]]></category>
            <category><![CDATA[ecommerce]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 19:42:12 GMT</pubDate>
            <atom:updated>2026-10-26T19:42:12.340Z</atom:updated>
            <content:encoded><![CDATA[<h4>Retailers evaluating augmented reality eCommerce vendors almost always start with the demo, and the demo is the wrong place to start.</h4><p>The question that decides the project is not whether a vendor’s AR looks good on stage. It is who digitizes the catalog, and at whose cost.</p><p>Several AR try-on plans assume a retailer already has AR-ready 3D models and price the model-creation work separately, off the plan you were quoted. That single line item has derailed more launches than any tracking-accuracy complaint we hear from buyers.</p><p>We build one of the virtual try-on engines compared below, so our selection criteria come before the vendor list, and every vendor gets an honest read on what it does best.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mHaW4qfNG--Svkdcyub6Og.png" /></figure><h3>Six checks to run before any demo</h3><p>Work through these in order. They are ranked by how often they kill a project after the contract is signed, not by how often a vendor brings them up first.</p><ol><li><strong>Who digitizes the catalog, and at whose cost?</strong> <br>This is the line item most pricing pages omit. Banuba generates virtual cosmetics from product parameters in an admin panel and builds eyewear frames from a couple of product photographs per frame, so digitization stays in-house instead of becoming a separate studio invoice.</li><li><strong>How many categories does one contract cover?</strong> <br>A retailer selling makeup, glasses, and hair color either buys one engine for all three or negotiates three separate ones.</li><li><strong>Web, app, or both?</strong> <br>A widget reaches every visitor who lands on the product page. A native SDK only reaches shoppers who already installed the app.</li><li><strong>Where does the camera frame go?</strong> <br>Processing on the shopper’s device keeps their image off a server, which shortens the privacy review before legal signs off.</li><li><strong>What does the storefront integration actually involve?</strong> <br>A packaged app for a platform like Shopify is a few clicks. Anything else is an engineering ticket with a timeline attached.</li><li><strong>Is pricing published?</strong> <br>A vendor that quotes only on request turns budgeting into an email thread with no anchor number.</li></ol><h3>Augmented reality eCommerce vendors, compared</h3><p>This is not a ranking. Each vendor is genuinely strong at something different, and the right one depends on which of the six checks above matters most for the catalog in front of you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NUx_5p_dQcLskIkY.jpg" /></figure><h4><strong>Banuba</strong></h4><p>Multi-category try-on behind one web widget, with self-service digitization and on-device processing. The self-serve plans cover makeup, skincare, eyewear, contact lenses and hair dye; an enterprise build adds headwear and jewelry.</p><h4><strong>GlamAR</strong></h4><p>Published per-SKU pricing tiers, with try-on across makeup, eyewear, jewelry, watches, footwear and apparel on iOS, Android and web.</p><h4><strong>Revieve</strong></h4><p>A modular JavaScript SDK that wraps skin diagnostics and product advice around the try-on step, rather than only the overlay.</p><h4><strong>ZERO10</strong></h4><p>In-store AR mirror installations plus a native iOS try-on framework, aimed at fashion retail.</p><h4><strong>Zappar Universal AR</strong></h4><p>Browser-first campaign AR triggered by packaging or print, with open documentation and no login required.</p><h4><strong>Snap Camera Kit</strong></h4><p>Official SDKs for iOS, Android and the modern browser, with the distinction that the audience is already AR-native.</p><h4><strong>MediaPipe and Google ML Kit</strong></h4><p>Free face-detection and mesh building blocks for a team building its own tracking from scratch. Neither ships a catalog, a shade renderer or an admin panel.</p><p>Two more names come up in this space without being try-on engines at all. <strong>Zakeke </strong>and <strong>Threekit </strong>sit in the <em>3D-configurator </em>and <em>product-visualization</em> layer.</p><p>Banuba and Zakeke work together on the eCommerce side, because a configurator and a try-on engine solve different problems rather than competing for the same one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*94tqWU5-VLcr_g_h.jpg" /></figure><h3>What this costs, and who will tell you</h3><p>Of the eight options above, only two publish a price at all: Banuba and GlamAR. Everyone else either quotes on request or has no pricing page.</p><p>Banuba runs two separate ladders, and it matters which one you are being quoted. The <a href="https://www.banuba.com/virtual-try-on#pricing">self-serve platform</a> lists $99 a month for 1,000 try-ons, $299 for 3,500 and $799 for 15,000, each with a 7-day trial and unlimited products. The Shopify app is priced for a different shape of traffic, from $249 a month for up to 2,000 try-ons to $1,299 for 40,000.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*Wesov8kf6ppO6mkh7qPmyg.png" /></figure><p>Those are list prices, and a promotional rate runs on the platform page from time to time, so check the live table before you budget against them. A custom enterprise integration, which is where full white-label branding and the headwear and jewelry categories live, is quoted on monthly try-on sessions and category count.</p><p>The planning consequence of the rest of the list is simple: most of your shortlist conversations will start without a number attached.</p><h3>Web widget or native SDK: how retailers actually decide</h3><p>Web wins whenever the goal is conversion on the product page, because it skips the app install, which is the single largest drop-off point in a shopping funnel.</p><p>Banuba’s web widget runs in the browser on nearly any connected device, so the shopper taps a button and the camera opens with no download in between.</p><p>The trade-off is that the browser sets the ceiling. Camera access needs HTTPS, and support tracks the last two major Safari and iOS releases plus the last five releases of Chrome, Edge, Opera and Samsung Internet.</p><p>A native SDK earns its engineering cost when try-on is one feature inside a longer session a retailer already owns, such as a loyalty app with repeat visits, or when the experience needs a device capability the browser does not expose.</p><p>Most retailers we work with start with the web widget and add an app-based path later, once the web version has already proven the category converts.</p><h3>What the evidence in this category actually says</h3><p>This is the part most vendor pages skip, and we are going to do it before making a claim of our own, because the honest picture changes how you should read every number in this market, including ours.</p><p><strong>There is almost no independently verifiable try-on performance data.</strong> Nearly every circulating conversion or return figure traces back to a vendor’s own marketing page. No eyewear retailer has published a try-on conversion or return figure in an earnings call or an audited release. Within a single competitor’s own case studies, published conversion uplift ranges from +9% to +200%, and most of those compare shoppers who chose to use the try-on against shoppers who did not, which measures self-selection rather than lift.</p><p>Several figures you will see quoted at you do not survive a check. The widely repeated “25 to 40% reduction in returns, per McKinsey” has no such McKinsey source. The “94% higher conversion with 3D and AR” number traces to a single Shopify social post from 2020 about interaction with 3D content, with no study, sample or method ever published, and it was silently restated two years later as a causal merchant-level claim.</p><p><strong>What does hold up is demand-side data and one piece of academic work.</strong></p><p>Research by Episto, fieldwork of 339 eyewear buyers in France and the US in January 2025, found that <a href="https://www.opticianonline.net/content/news/two-thirds-of-online-buyers-use-virtual-try-on-says-survey">29% of glasses shoppers have used a virtual try-on, rising to 67% among those who buy online</a>. The study was commissioned by a vendor in this category, which is worth knowing, though the fieldwork house is independent.</p><p>For the returns conversation, the right denominator is <a href="https://nrf.com/research/2025-retail-returns-landscape">the NRF’s 2025 figure that 19.3% of online sales come back</a>. That is independent and current, and it is context rather than a claim about what try-on does to it.</p><p>The strongest quantitative work is academic. Tan, Chandukala and Reddy, writing in the <em>Journal of Marketing</em> in 2022 with real cosmetics sales data, found that AR usage is associated with higher sales for less popular brands, narrow-appeal products and more expensive ones. They describe the overall sales effect as small.</p><p>That finding is more useful to a buyer than any vendor percentage, because it tells you where the effect concentrates: the long tail of your catalog, not the hero products that already sell.</p><h3>Why Banuba</h3><p>Our own figure, and it is ours rather than independent. Océane, a Brazilian beauty brand, took its add-to-cart rate from 3% to 32% after adding Banuba’s virtual try-on widget. That is a makeup deployment, so read it as evidence about the engine rather than as a benchmark for eyewear or any other category.</p><p>Beyond that, the argument is a cost argument rather than a performance one, which is the part a retailer can actually verify before signing.</p><p>More than 22,000 products are already digitized on the platform, and a retailer can add a new collection in under 48 hours, because the frame and shade data comes from photographs and product parameters rather than a 3D modeling queue.</p><p>Cosmetics are generated from catalog parameters in an admin panel. Eyewear needs a couple of product photographs per frame, the kind most stores already have on their product pages.</p><p>That is the line item from check one, and it is the one we would rather be judged on than on a conversion percentage nobody in this category can substantiate.</p><h3>FAQ</h3><h4><strong>Is augmented reality worth adding to an eCommerce store?</strong></h4><p>For categories where fit or shade drives returns, meaning makeup, eyewear and hair color, the case is reasonable, but be careful which evidence you accept for it.</p><p>Demand-side data is solid: 67% of people who buy glasses online have already used a virtual try-on somewhere. Performance data is weak across the whole category, ours included, and the one peer-reviewed study with real sales data describes the overall effect as small while finding it concentrated in less popular and more expensive products.</p><p>The honest way to decide is a trial on your own catalog rather than a percentage from a vendor page.</p><h4><strong>How much does an AR try-on widget cost?</strong></h4><p>Banuba’s self-serve platform lists $99 a month for 1,000 try-ons and scales to $799 a month for 15,000. A Shopify-specific plan starts at $249 a month for up to 2,000 try-ons.</p><p>A custom enterprise integration with full white-label branding is quoted separately, based on monthly try-on sessions and category count. Current figures are on the <a href="https://www.banuba.com/virtual-try-on#pricing">pricing table</a>.</p><h4><strong>Does a retailer need a 3D model of every product before launching AR try-on?</strong></h4><p>Not with Banuba. Cosmetics are generated from product parameters entered in an admin panel, and eyewear frames are built from a couple of product photographs each, so a retailer can digitize a catalog without a dedicated 3D studio or CAD pipeline.</p><p>An enterprise build adds a higher-precision 3D route for catalogs where the frame has to render exactly.</p><h4><strong>Can shoppers use AR try-on without downloading an app?</strong></h4><p>Yes. Banuba’s try-on opens from the product page, the shopper taps the try-on button and the browser opens the camera, with no install step.</p><p>An app-based SDK is only worth the engineering cost once try-on is one feature inside a session a retailer already owns.</p><h4><strong>Which categories can one contract cover?</strong></h4><p>Banuba’s self-serve plans cover makeup, skincare, eyewear, contact lenses and hair dye on every tier, with the tiers differing by monthly try-on volume rather than by feature.</p><p>Headwear and jewelry sit on an enterprise build, alongside personalized AI recommendations and priority support.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ade32f4928f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 Virtual Eyewear Try-On Platforms Compared: Catalog, Integration, and Pricing (2026)]]></title>
            <link>https://banuba.medium.com/virtual-eyewear-try-on-compared-0e3e9737740f?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e3e9737740f</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[shopify-development]]></category>
            <category><![CDATA[ecommerce]]></category>
            <category><![CDATA[augmented-reality]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 19:26:18 GMT</pubDate>
            <atom:updated>2026-10-26T19:26:18.697Z</atom:updated>
            <content:encoded><![CDATA[<p>In this guide, we compare the five vendors that come up on every virtual eyewear try-on shortlist: Fittingbox, Auglio, Banuba, Perfect Corp, and GlassOn.</p><p>The question that actually separates them is who owns the frames. A vendor with an existing 3D database saves a multi-brand retailer weeks of digitization work, while a brand digitizing its own collection is paying for coverage it cannot use.</p><p>All five render try-on through the browser camera in real time. That part of the category has stopped being a differentiator. What still separates them is catalog economics, integration path, and how openly each one prices.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LXDg9oO61yOlT82i2YEo6w.png" /></figure><h3>The four questions that actually separate these vendors</h3><h4><strong>Who owns the frames</strong></h4><p>If you sell other brands’ frames, a vendor with an existing 3D database saves the digitization work entirely. If you sell your own collection, that database is worth nothing to you.</p><h4><strong>How the catalog gets digitized</strong></h4><p>All five work from product photographs rather than physical samples, so the question is how many photographs per frame and how fast the model comes back. The difference is measured in lead time on every new collection you add.</p><h4><strong>Where it has to run</strong></h4><p>All five run in the browser. What varies is whether a ready-made storefront app exists for your platform, or whether a developer has to wire up the integration by hand.</p><h4><strong>How you are billed</strong></h4><p>Try-on volume, SKU count, seats, and flat license fees are all in use across this set, and three of the five vendors will not give you a number until you talk to sales.</p><h3>Virtual eyewear try-on software compared</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zPuWXjovweveYnU7DfAGqg.png" /></figure><p>Each of these is a different shape of product: an SDK, a packaged app, a plugin, or a web component set.</p><p>So treat the table as a starting shortlist rather than a like-for-like spec sheet. Fittingbox and Auglio are catalog-and-integration plays, Banuba and Perfect Corp both ship a metered Shopify app on top of a broader platform, and GlassOn is the smallest vendor of the five, built specifically around the lens.</p><p>Read per criterion, the winners are:</p><h4><strong>Frame database size goes to Fittingbox</strong></h4><p>At 195,000-plus references across 1,200-plus brands, <a href="https://fittingbox.com/en">it is the only vendor publishing a number here</a> at all, and none of the other four have anything comparable on the public record.</p><h4><strong>Raw pre-virtualized catalog breadth goes to Auglio</strong></h4><p>At <a href="https://auglio.com/">500,000-plus products already built</a>, if your frames happen to already be inside it.</p><h4><strong>Digitization input is a three-way tie at the top</strong></h4><p>Auglio works from one 2D photo, GlassOn from one or two product photos, and Banuba from a couple of product photographs per frame. None of the three asks for physical samples, so for a brand with product photography already on its detail pages, the asset step is not the thing that separates them.</p><h4><strong>Speed for a brand digitizing its own collection goes to Banuba</strong></h4><p>New collections go live in under 48 hours because the try-on does not depend on any third-party frame database, only on what you upload yourself.</p><h4><strong>Lens visualization goes to GlassOn outright</strong></h4><p>It is the only vendor in this set that treats lens tint and gradient comparison as a shipped feature rather than a rendering setting, and nothing else here matches it.</p><h4><strong>Category breadth on a single contract goes to Banuba</strong></h4><p>Covering eyewear alongside makeup, skincare, contact lenses and hair dye on one self-serve plan. Perfect Corp is the next broadest, with its Shopify app covering up to 10 makeup and eyewear categories on one contract.</p><h4><strong>Published entry pricing goes to Banuba</strong></h4><p>At $99 a month for the self-serve widget against no comparable published number from three of the other four vendors, and against Perfect Corp’s $379-a-month Shopify entry tier as the one direct comparison available.</p><h3>What each vendor is actually built for</h3><p><strong>Fittingbox</strong> has been in the category for <a href="https://fittingbox.com/en/about-us">20 years and holds 59 international patents</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mQemEr7WN6Z7LodRXkizoA.png" /></figure><p>Its case for a multi-brand optical retailer is that the frames on the shelf are probably already in its database, so the slowest part of a rollout is already done. It also ships <a href="https://fittingbox.com/en/optical-fit/online-pd-measurement-tool">pupillary-distance measurement accurate to within 1 millimeter on 7 of 10 measurements</a> and a frame-removal feature, so a shopper can try frames without taking off the pair they are already wearing.</p><p>The trade-off: no published pricing, and the database is only worth what you actually use of it.</p><p><strong>Auglio</strong> runs an eyewear-first plugin on top of its 500,000-plus-product catalog and states that a try-on needs nothing more than a 2D product photo.</p><p>Around the core try-on it adds automatic PD measurement, an invisible-glasses mode, 360-degree frame viewing, photochromic lens visualization, and face shape detection. Its public client list includes JBL, Bupa Optical, and Goodr, and it has published a case study for the retailer Zoff reporting a fourfold increase in conversions.</p><p>The trade-off: pricing is not published, and its eyewear, cosmetics, and wig offerings are split across separate domains.</p><p><a href="https://www.banuba.com/glasses-virtual-try-on"><strong>Banuba</strong></a> is built for brands that own their frame collection and do not want to wait on a third-party database.</p><p>A frame goes into the Banuba admin dashboard as a couple of product photographs plus its name, brand, and SKU, the AI builds the model from them, and the preview renders as you upload. For catalogs where the frame has to render exactly, an enterprise build adds a higher-precision 3D route that captures each frame from three angles.</p><p>The widget <a href="https://apps.shopify.com/tint-virtual-try-on-by-banuba">ships as a script tag or a Shopify app</a> and runs on the last two Safari versions and the last five Chrome, Edge, Opera, and Samsung Internet versions. A <a href="https://dev.to/team_banuba/virtual-eyewear-try-on-4-jobs-between-frame-photos-and-cart-8ep">developer walkthrough of the frame-photo-to-cart integration steps</a> covers that path in more technical detail.</p><p>Pupillary-distance measurement is included with no special hardware, and eyewear sits on the same self-serve plan as makeup, skincare, contact lenses and hair dye, so a brand selling more than frames is not buying a second system. You can <a href="https://demo.tintvto.com/">test the widget yourself</a> before talking to anyone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*C26j7pWL5GlKSX6_VF9N_w.png" /></figure><p>The clearest published number behind the engine is not from eyewear. The cosmetics brand Océane <a href="https://www.banuba.com/blog/banuba-helped-oceane-achieve-a-record">reached a peak add-to-cart rate of 32%</a> after adding a makeup try-on, against the 3% baseline Banuba cites for the category. That is our own published figure from a single makeup deployment, so read it as evidence about the engine rather than as an eyewear benchmark.</p><blockquote>The trade-off: if you sell frames you do not own, you are digitizing a catalog someone else has already digitized.</blockquote><p><strong>Perfect Corp</strong> sells the <a href="https://www.perfectcorp.com/business/services/sdk">YouCam SDK for Mobile and YouCam for Web Enterprise</a>, plus a no-code Shopify app.</p><p>Its real-time tracking is patented under the name AgileFace, its facial analysis model is trained on more than 100,000 faces and detects more than 70 facial traits, and its web try-on runs on edge computing, so shopper data does not have to leave the device.</p><p>The Shopify app is the only competitor here besides Banuba that publishes a number: <a href="https://apps.shopify.com/youcam-makeup-official">$379 a month for 100 SKUs on its Essential tier, $569 a month for 300 SKUs on Premium</a>, each with a 14-day trial. SDK licensing itself is contact-sales only.</p><p>The trade-off: the published entry price is the highest in this set, and the SKU allowance is tight for a full frame catalog.</p><p><strong>GlassOn</strong> is the only vendor that gives the lens equal billing with the frame, on the logic that lenses are roughly half the value of a pair of glasses but entirely invisible to an online shopper.</p><p>Alongside a fit-accurate try-on, it ships tint and gradient comparison, side-by-side lens comparison, and the ability to try new lens colors on the glasses a shopper is already wearing. It <a href="https://glasson.io/">builds 3D models from one or two product photos</a>, offers 360-degree rotation and physical-glasses removal, and integrates through a Shopify app, a PrestaShop add-on, or web components.</p><p>The trade-off: pricing is quote-only, and it is the smallest vendor on this list.</p><h3>What it costs, and why the two metering models do not compare directly</h3><p>Only Banuba and Perfect Corp publish a price at all, which is the single most useful fact to have before a procurement round starts.</p><p>Banuba runs two separate price ladders, and it matters which one you are being quoted.</p><p>The <a href="https://www.banuba.com/virtual-try-on#pricing">self-serve platform</a> lists $99 a month for 1,000 try-ons on Basic, $299 for 3,500 on Pro, and $799 for 15,000 on Smart, plus a custom Enterprise tier. Each comes with a 7-day trial, unlimited products, and makeup, skincare, eyewear, contact lenses and hair dye on every tier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HjnL2A7YQgxIHOcj.png" /></figure><p>The Shopify app is priced for a different shape of traffic: $249 a month for up to 2,000 try-ons on Basic, $799 for 10,000 on Pro, and $1,299 for 40,000 on Smart, each with a 7-day trial and unlimited products, covering makeup, skincare, glasses and hair dye.</p><p>Those are list prices. A promotional rate runs on the platform page from time to time, so check the live table before you budget against the figures above.</p><p>Perfect Corp’s Shopify app publishes $379 a month on Essential and $569 a month on Premium, both with a 14-day trial, but caps SKUs at 100 and 300 respectively rather than metering by try-on volume.</p><p>That difference in metering matters more than the headline numbers do. Banuba charges by try-on volume; Perfect Corp charges by SKU count.</p><p>A large catalog with modest traffic and a small catalog with heavy traffic get opposite answers out of those two models, so run both against your own numbers before comparing the sticker price.</p><p>Fittingbox, Auglio, and GlassOn all route to a contact form, which is not a criticism, just a planning fact: three of your five shortlist conversations start without a number attached.</p><p>Every competitor price above is what each vendor published in October 2026. Vendor pricing pages change, ours included, so confirm current numbers directly before you build a budget around them.</p><p>None of the three quote-only vendors publishes whether that first sales call is a formality or a qualification gate. Banuba’s and Perfect Corp’s Shopify apps are the only two paths in this set you can start on your own card, without waiting on anyone.</p><h3>Where Banuba’s virtual eyewear try-on fits, and where it does not</h3><p>Two objections about Banuba come up often enough in this category to answer directly rather than leave for a footnote.</p><p><strong>The first is that Banuba is expensive.</strong> On published entry pricing, it is the opposite: the self-serve platform starts at $99 a month, and Shopify-app to Shopify-app, Banuba’s $249 entry tier is below Perfect Corp’s $379, while the other three vendors publish nothing to compare against at all.</p><p>Where the claim can become true is at high try-on volume, where the Shopify Smart tier reaches $1,299 a month and the platform Smart tier $799, or on a custom Enterprise build. So the honest comparison asks for the volume curve rather than the headline number.</p><p><strong>The second is that Banuba is general-purpose rather than eyewear-first</strong>, and that one is fair.</p><p>It cuts both ways. Fittingbox and GlassOn are eyewear specialists, and that shows up in features Banuba does not have, including a 195,000-frame database and a dedicated lens-comparison mode. What a cross-category vendor buys instead is one contract, one integration, and one dashboard across eyewear, makeup, skincare, contact lenses and hair dye.</p><p>If frames are all you sell, that specialization gap should count against Banuba. If frames are one category among several, it is the reason to weigh it the other way.</p><p>Three situations where a different vendor on this list is the better answer:</p><ul><li><strong>You sell frames you do not own.</strong> Fittingbox’s database removes digitization work you would otherwise be doing yourself.</li><li><strong>Lens upgrades are a meaningful revenue line for you.</strong> GlassOn’s tint and gradient comparison is a shipped feature, and nothing else here matches it.</li><li><strong>Eyewear is a small module inside a larger beauty program you already run on Perfect Corp.</strong> Consolidating on one vendor can beat a marginally better standalone eyewear experience.</li></ul><p>Outside those three cases, Banuba’s narrow claim is that a brand digitizing its own frame collection can be live without waiting on anyone else to model that collection first, on a plan that also covers the rest of what it sells.</p><p>For a longer practitioner walkthrough of how to weigh these criteria against your own catalog, <a href="https://medium.com/@fedulovelena/how-to-choose-eyewear-virtual-try-on-a2421e72c0e6">this breakdown of the eyewear try-on buying decision</a> is worth reading alongside this comparison.</p><h3>Which vendor fits comes down to who owns the frames</h3><p>If your frames belong to other brands, Fittingbox’s database is the deciding factor, and no amount of rendering polish elsewhere changes that.</p><p>If lens choice drives revenue for you, GlassOn is the only vendor built around it.</p><p>If eyewear is one line in a beauty program you already run on Perfect Corp, staying put is a reasonable call.</p><p>If you want fast coverage on frames already sitting in a large third-party catalog, Auglio gets you there.</p><p>And if you own your collection and want to be live on it without depending on anyone else’s frame database, that is the case Banuba is built for, and the 7-day trial on either its platform or its Shopify app costs nothing to test.</p><h3>FAQ</h3><h4><strong>What is Banuba’s virtual eyewear try-on software, and how does it compare to the rest of this list?</strong></h4><p>It is a browser-based widget that renders a shopper’s own frames on their face through the device camera, with no app install required, so the shopper can judge fit before buying.</p><p>Unlike Fittingbox and Auglio, which draw on their own pre-built frame databases, Banuba’s version is built for brands uploading and digitizing their own collection.</p><h4><strong>Do I have to ship physical frames to Banuba to get them digitized?</strong></h4><p>No. You upload a couple of product photographs of each frame in the Banuba admin dashboard, along with the name, brand, and SKU, and new collections are typically live in under 48 hours.</p><p>Auglio and GlassOn also work from product photos rather than physical samples. Fittingbox’s advantage is that many mainstream frames may already be in its database and need no digitization at all.</p><h4><strong>Can I add Banuba’s eyewear try-on to Shopify without a developer?</strong></h4><p>Yes. Banuba, Perfect Corp, and GlassOn all publish Shopify apps, and Fittingbox lists Shopify among its supported platforms. Banuba’s <a href="https://www.banuba.com/ar-virtual-try-on-plugin">Shopify plugin</a> is the no-code route if your store already runs on Shopify.</p><h4><strong>Does Banuba’s eyewear try-on work on older phones and browsers?</strong></h4><p>Banuba states its glasses try-on works across low-end and current-generation devices, and the widget targets the last two Safari and iOS major versions and the last five Chrome, Opera, Edge, and Samsung Internet versions.</p><p>For all five vendors in this comparison, browser support is the practical constraint rather than raw device power.</p><h4><strong>Can Banuba cover eyewear and makeup on the same contract?</strong></h4><p>Yes, and it is the main argument for choosing a cross-category vendor over an eyewear specialist. Banuba’s <a href="https://www.banuba.com/tint-makeup-virtual-try-on">makeup virtual try-on</a> sits on the same self-serve plans as eyewear, skincare, contact lenses and hair dye.</p><p>Perfect Corp’s Shopify app similarly covers up to 10 makeup and eyewear categories.</p><h4><strong>What conversion lift should I expect from adding Banuba’s try-on to an eyewear store?</strong></h4><p>Be skeptical of any single number here, including ours. Published results range from Auglio’s Zoff case study reporting a fourfold conversion increase to Banuba’s cosmetics customer Océane reaching a 32% add-to-cart rate in makeup.</p><p>Both are real, both are single deployments in different categories, and neither predicts what happens on your own catalog. The trial exists to answer that question on your own numbers.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e3e9737740f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[6 steps to face verification with an API, from capture to threshold]]></title>
            <link>https://banuba.medium.com/face-verification-api-steps-e4952b8baa1f?source=rss-e1a005e0267b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e4952b8baa1f</guid>
            <category><![CDATA[api]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[face-recognition]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[biometrics]]></category>
            <dc:creator><![CDATA[Banuba]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 14:34:48 GMT</pubDate>
            <atom:updated>2026-10-23T14:34:48.160Z</atom:updated>
            <content:encoded><![CDATA[<h4>Face verification is a 1:1 check that answers a single question: is the person in front of the camera the same person as the enrolled template?</h4><p>Adding it to an app comes down to six decisions, in order: confirm the job is 1:1, choose where matching runs, standardize capture, set a threshold, add liveness, and protect the biometric data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Uty-tpg7-Ysaf_7d-WnH9g.png" /></figure><p>We build Banuba Face API, one of the tools covered below, so each step says where it fits and where it does not.</p><p>Before you start, have these four things settled:</p><ul><li>One reference image or stored template per user, captured at enrollment.</li><li>A camera stream at a consistent resolution.</li><li>A decision on whether frames may leave the device.</li><li>A written answer to what one wrong accept costs your business.</li></ul><h3>Step 1: Confirm the job is 1:1, not 1:N</h3><p>Face verification is 1:1. The app holds one enrolled template and asks whether the live face matches it. Login, re-authentication, and most KYC checks work this way.</p><p>Face identification is 1:N. The app searches a probe face against a database of enrolled identities to answer who the person is. Watchlists, access control at scale, and media indexing need this.</p><p>Face detection and face tracking are separate jobs again. Detection finds a face in a frame, and tracking follows it across frames and returns landmarks and head pose.</p><p>Most buying mistakes on this topic come from treating these capabilities as one product.</p><p>The choice narrows the vendor list immediately. Google Cloud Vision detects faces and localizes landmarks, but it has no face matching, no enrollment, and no identity search, so it cannot be the matching engine for either flow.</p><h3>Step 2: Should face matching run in the cloud or on the device?</h3><p>In the cloud when you need identity search or an audited certification; on the device when frames must not leave the handset. That is the short version, and the rest of this step is what each option costs you to get there.</p><p>A cloud API is reachable from any language over REST, which makes it stack-agnostic and network-dependent. An on-device SDK is the opposite trade: you integrate per platform, and in exchange the camera pipeline never leaves the handset.</p><p>Here is what each option offers for a 1:1 flow, as we read the vendor pages in September 2026.</p><h4><strong>Amazon Rekognition</strong></h4><p>Cloud only. It covers 1:1 face comparison and 1:N search against collections you index.</p><p>Pricing starts at <a href="https://aws.amazon.com/rekognition/pricing/">$0.0010 per image for the first 1 million images per month</a>, and the free tier lasts 12 months from account creation and covers 1,000 images per month.</p><h4><strong>Microsoft Azure Face</strong></h4><p>Cloud only. Face Verification is a Limited Access feature: you apply through Microsoft’s Face Recognition intake form, and it runs on the Standard (S0) and Enterprise (E0) tiers, not on the free F0 tier, as the <a href="https://learn.microsoft.com/en-us/azure/ai-services/face/overview-identity">Azure Face identity overview</a> describes it.</p><p>Billing runs per 1,000 transactions on the <a href="https://azure.microsoft.com/en-us/pricing/details/cognitive-services/face-api/">Azure Face pricing page</a>, with Liveness and Liveness with Verification on separate lines from verification itself. We could not get that page to render figures for a region, so we are not quoting rates for it here. Pull yours from the pricing calculator.</p><p>Microsoft also states that face recognition features may not be used by or for U.S. police departments.</p><h4><strong>Face++</strong></h4><p>Cloud APIs with SDKs for web, Android, and iOS, including 1:1 comparison. Self-serve sign-up was open when we checked.</p><p>The <a href="https://www.faceplusplus.com/v2/pricing/">pricing page</a> publishes no numeric prices, only four price models: Free, Pay As You Go, a Daily or Monthly QPS capacity plan, and Pay for Licensing.</p><p>One procurement step applies here and not to the others. If you buy in a regulated, government-adjacent or defense-adjacent context, run Face++ and its parent company through your own restricted-party screening before you spend time evaluating. That check belongs to your compliance team rather than to a vendor comparison, and it is faster to run it first than to unwind an evaluation later.</p><h4><strong>Google Cloud Vision</strong></h4><p>Cloud only, and no face matching, as noted in Step 1.</p><p><a href="https://cloud.google.com/vision/pricing">Face Detection costs $1.50 per 1,000 units</a> for units 1,001 to 5,000,000 per month, and the first 1,000 units per month are free per feature.</p><h4><a href="https://www.banuba.com/face-api"><strong>Banuba Face API</strong></a></h4><p>It runs on the device by default, so no frame is uploaded and there is no per-request charge. Licensing is MAU-based through Banuba Face AR SDK, with a 14-day free trial.</p><p>It supplies 1:1 face match against one reference image you provide, plus the signals for a liveness challenge. It is a tracking and effects product rather than an identity product: enrollment, decision logic and any KYC layer are yours to build, and it does not run identity search against an enrolled database.</p><p>The cloud platforms win on identity search, because Rekognition and Azure both support 1:N and Banuba Face API does not. Azure also wins on published liveness evidence, covered in Step 5.</p><p>Banuba has not submitted its models to NIST FRVT or published LFW or MegaFace results, so any accuracy comparison against the cloud vendors is untested rather than won.</p><p>Vendor pages change, and the prices above are approximate, not quotes. A figure that decides your build should come from the vendor in writing.</p><h3>What does a per-transaction bill look like?</h3><p>It is not the verification line that decides this. On every cloud vendor we looked at, liveness is billed separately from the match itself and it is the expensive one, so the number that moves your architecture is the liveness rate rather than the verification rate.</p><p>We are not putting a figure on it, because we could not get Azure’s pricing page to render rates for a region and we are not going to quote a price we did not read. Pull both rates for your own region first, then run four lines:</p><ol><li>Daily verifications x 30 = monthly volume.</li><li>Monthly volume / 1,000 = billable thousands.</li><li>Billable thousands x your verification rate = the cost of matching.</li><li>Billable thousands x your liveness-with-verification rate = the cost of the flow.</li></ol><p>Line 4 is the one to take into the room. A licensed on-device SDK puts the same volume into a different cost model entirely, where request count is not a variable, so the comparison worth making is line 4 against a licence rather than line 3 against anything.</p><p>Run it before the pilot rather than after, because this is the calculation that changes which architecture you pick, and it changes it at a volume most teams reach in their first year.</p><h3>Step 3: Standardize capture before you tune anything</h3><p>Most accuracy problems in a verification flow start at capture.</p><p>The failure modes repeat across implementations: low-resolution crops and over-compressed frames, poor alignment from wrong bounding boxes or inconsistent landmarking, enrolling in ideal light and verifying in the dark, and multi-face scenes where the wrong person gets matched.</p><p>Write three capture rules before touching any threshold: a minimum face size, lighting guidance on the capture screen, and identical preprocessing at enrollment and at matching. In frames with more than one face, choose the face to verify deliberately instead of letting the matcher pick.</p><p>Alignment is the part teams underestimate. Banuba’s product page states 68 facial anchor points per face, and a consistent landmark set is what keeps alignment stable between enrollment and matching.</p><p>Detection quality upstream sets the ceiling for everything downstream, and an open-source detector is a reasonable start for a prototype. Our <a href="https://banuba.medium.com/banuba-face-ar-sdk-vs-opencv-a-technical-comparison-1ba68df81c3a">comparison of OpenCV and Banuba face detection</a> shows where that stops being cheaper once your team owns the maintenance.</p><h3>Step 4: How do you set the face matching threshold?</h3><p>Start from your risk profile. Security and fintech flows prioritize low false accepts, which means a stricter threshold and a slightly higher false reject rate. Consumer and AR apps prioritize low false rejects, because every rejected user is friction.</p><p>Then tune on your own data. Collect a few hundred representative captures across the lighting, camera hardware, pose, and demographics you expect, and measure false accepts and false rejects separately for each segment.</p><p>A single threshold for everyone is usually a production incident waiting to happen. Public leaderboard datasets such as LFW and MegaFace were built for research comparison and say little about a user holding a phone in a dim room.</p><p>Here is what a false accept rate means at volume. This example runs at a different scale from the billing one above, deliberately, because the threshold question only gets interesting at scale. Both rates are illustrative, so replace them with the ones you measure.</p><ol><li>Daily verifications: 100,000.</li><li>False accept rate at a looser threshold: 0.1%, or 0.001.</li><li>Wrong accepts per day: 100,000 x 0.001 = 100.</li><li>False accept rate at a stricter threshold: 0.01%, or 0.0001.</li><li>Wrong accepts per day: 100,000 x 0.0001 = 10.</li><li>Over a 30-day month, the looser threshold lets through 100 x 30 = 3,000 wrong accepts, and the stricter one lets through 300.</li></ol><p>Multiply the monthly count by the answer you wrote down before you started: what one wrong accept costs. That number is what sets the threshold, and a vendor default is a starting point rather than a decision.</p><h3>Step 5: Add liveness, and ask for the evidence</h3><p>A matcher compares faces. It does not know whether the face is a live person, a photo, or a screen replay, so a verification flow needs a liveness check in front of it.</p><p>Ask every vendor for accredited presentation attack detection results rather than a claim: the test level, the laboratory, and the standard. The usual answer is iBeta Level 1 or Level 2, conformant to ISO/IEC 30107–3.</p><p>Microsoft reports that its liveness detection, itself a gated feature, achieved a 0% penetration rate in iBeta Level 1 and Level 2 testing by a NIST/NVLAP-accredited laboratory. That is Microsoft’s published result, and Microsoft sells the service, so read it as evidence to ask other vendors for, not as our finding.</p><p>Banuba Face API takes the active route. It tracks head pose, eye openness, mouth movement, gaze, and pulse, and supports blink, smile, and head-turn triggers, so your app issues a challenge and confirms the response.</p><p>The decision logic is yours to build, and Banuba is not the pick when you need an audited presentation attack detection certification. If a regulator or auditor asks for one, choose one of the cloud vendors. Our <a href="https://www.banuba.com/blog/liveness-detection">liveness detection guide</a> explains how active and passive checks differ.</p><p>Plan the surrounding workflow as well: a fallback path for poor lighting, older devices, and accessibility needs, and periodic re-validation as new spoof patterns appear.</p><h3>Step 6: Handle the biometric data deliberately</h3><p>Most systems store a face template or embedding instead of a raw image, but an embedding is still a biometric identifier, so the regulatory risk stays.</p><p>Store the minimum you need, set retention windows and delete on request, and encrypt templates at rest and in transit. Keep enrollment, matching, and user identifiers separated under least-privilege access.</p><p>Where regulations, latency, or data residency are strict, prefer on-device or on-premises processing. On-device is the strongest structural control, since frames never leave the handset, and it is how Banuba Face API runs by default.</p><h3>Most of a face verification flow is decided before the API call</h3><p>The API call is one line of the integration. The job definition, the place matching runs, capture rules, the threshold, and the liveness evidence decide whether the flow works in production.</p><p>If the project needs identity search or an audited presentation attack detection certification, start with Amazon Rekognition or Microsoft Azure Face, and plan for the approval step Azure requires.</p><p>If it needs 1:1 match and liveness signals inside a mobile or web app where frames stay on the device, Banuba Face API supplies the signals and your team owns the decision.</p><h3>FAQ</h3><h4><strong>What is the difference between face verification, face identification and face recognition, and which does Banuba Face API cover?</strong></h4><p>Face verification is a 1:1 check that answers “is this person who they claim to be?” Face identification is a 1:N search against a database of enrolled identities that answers “who is this person?” Face recognition is the umbrella term for both, plus enrollment and matching.</p><p>Banuba Face API supplies 1:1 face match against a reference image you provide, and it is not an identity search product.</p><h4><strong>How should I set the matching threshold when I build face verification on Banuba Face API?</strong></h4><p>Set it from your risk profile, then tune it on your own captures. Security and fintech flows should favor low false accepts, and consumer and AR apps should favor low false rejects.</p><p>Validate each segment separately, because lighting, camera quality, pose, and demographics shift the curve.</p><h4><strong>Does Banuba Face API handle liveness detection for face verification?</strong></h4><p>Banuba Face API gives you the inputs for a liveness check rather than a finished verification product. It tracks head pose, eye openness, mouth movement, gaze, and pulse, and supports active triggers such as blink, smile, or head turn.</p><p>The decision logic and any KYC or compliance layer are yours to build.</p><h4><strong>Does Banuba Face API send face data to the cloud?</strong></h4><p>By default, no. Banuba Face API runs entirely on the device: the camera frame is processed on the phone, browser, or desktop and is not uploaded anywhere, which is also why there is no per-request pricing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e4952b8baa1f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>