<?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 Obsidian Studios on Medium]]></title>
        <description><![CDATA[Stories by Obsidian Studios on Medium]]></description>
        <link>https://medium.com/@obsidianstudios?source=rss-621e7399d7a3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*gJ5uOPkvBQfbFGTGXCMf8A.png</url>
            <title>Stories by Obsidian Studios on Medium</title>
            <link>https://medium.com/@obsidianstudios?source=rss-621e7399d7a3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 11:25:25 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@obsidianstudios/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[Why most AI-generated ads still look fake (and it’s not the AI’s fault)]]></title>
            <link>https://medium.com/@obsidianstudios/why-most-ai-generated-ads-still-look-fake-and-its-not-the-ai-s-fault-8943ff61c6e1?source=rss-621e7399d7a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/8943ff61c6e1</guid>
            <category><![CDATA[video-production]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Obsidian Studios]]></dc:creator>
            <pubDate>Sun, 22 Nov 2026 10:36:41 GMT</pubDate>
            <atom:updated>2026-11-22T10:36:41.261Z</atom:updated>
            <content:encoded><![CDATA[<p>I spent an afternoon last month scrolling through a folder of AI-generated ad clips a client had made with four different tools. Within about ninety seconds I could tell which tool made which clip, and not because the visuals were bad. The problem was almost always the same: the character’s face wandered between shots. A woman holding a coffee cup in frame one had slightly different cheekbones by frame three. Nobody consciously catches that on a first watch, but something in the brain flags it anyway, the same way you notice a mirror reflection lagging a beat behind.</p><p>This is the actual bottleneck in AI ad production right now, and it has nothing to do with resolution or motion smoothness. It’s identity drift: the inability to keep a generated character looking like the same person across multiple shots, lighting setups, and angles. Most of the “AI killed video production” takes skip right past this, because a single pretty shot is easy and a consistent character across fifteen shots is hard.</p><p>A few things actually help, with real trade-offs attached. Locking a reference image and feeding it into every generation reduces drift but doesn’t erase it, since the model still reinterprets lighting and angle each time and small features drift anyway. Training a dedicated model on one face works better but costs real time and money, which is overkill for a fifteen-second product ad and underkill for a campaign meant to run all year. What’s actually held up for us, after testing most of the alternatives, is keeping shot count low and camera movement deliberate: fewer cuts, slower moves, and composition choices that hide the seams instead of fighting them. Less flashy than a five-shot montage, and also the difference between an ad that reads as competent and one that quietly tips a viewer off that something’s synthetic.</p><p>None of this means AI video is a dead end for advertising. It means the easy 80 percent of the job, getting one decent-looking shot, was never the hard part. The hard part was always continuity, whether a human shot it on 35mm or a model generated it frame by frame. A few shops, including ours at Obsidian Studios, have quietly rebuilt their pipeline around that one constraint instead of chasing whatever model shipped last week.</p><p>If you’re evaluating an AI ad vendor, ask to see the same character across at least four different shots before you ask about pricing. That one request tells you more about whether the work will hold up than any demo reel will.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8943ff61c6e1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Indian festive and regional ads with AI: what works and what breaks.]]></title>
            <link>https://medium.com/@obsidianstudios/indian-festive-and-regional-ads-with-ai-what-works-and-what-breaks-d43d02dc819a?source=rss-621e7399d7a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/d43d02dc819a</guid>
            <category><![CDATA[india]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[advertising]]></category>
            <dc:creator><![CDATA[Obsidian Studios]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 21:55:59 GMT</pubDate>
            <atom:updated>2026-11-19T21:55:59.984Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*id6mK8qRZdFaa76m.png" /></figure><p>AI video models learned mostly from global footage. Ask for Diwali, a Punjabi wedding or a Kerala home and you often get a generic, slightly foreign version of it. Indian audiences notice immediately.</p><h3>Festivals become decoration</h3><p>Models tend to reduce a festival to its most common visual: lots of diyas, lots of marigolds, everyone in new clothes. Real festive moments are specific to a family, a region and a time of day. Brief the moment, not the festival: who is in the room, what they are doing, what happens next.</p><h3>Weddings flatten into one style</h3><p>A Marwari, Bengali, Tamil and Punjabi wedding look very different, from clothing to jewellery to the rituals shown. Generic prompts blend them. Reference images from the right community, checked by someone who knows it, matter more than a longer prompt.</p><h3>Skin tones drift lighter</h3><p>Generated faces often drift towards lighter skin across a sequence, even when the first shot is right. We set skin tone in the character reference and check it across the whole edit and in the grade.</p><h3>Text in Indian scripts breaks</h3><p>Devanagari, Tamil, Bengali and other scripts are rarely rendered correctly inside generated frames. Any on-screen text in Indian languages should be added in the edit as real typography, never generated.</p><h3>Language versions need planning</h3><p>A film made in Hindi and later dubbed into Tamil feels dubbed. When a campaign will run in several languages, plan the script, pacing and supers for each version from the start.</p><h3>What works well</h3><p>Settings that need no cultural precision, such as product hero shots, abstract worlds, stylised animation and visual metaphors, are where AI shines for Indian brands. Combine those with carefully directed human moments and the result can feel both fresh and true.</p><p><em>Originally published at </em><a href="https://obsidianstudios.in/intelligence/festive-regional-ai-ads-india"><em>https://obsidianstudios.in</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d43d02dc819a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What actually goes into the cost of an AI ad film.]]></title>
            <link>https://medium.com/@obsidianstudios/what-actually-goes-into-the-cost-of-an-ai-ad-film-b36cf93213e2?source=rss-621e7399d7a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/b36cf93213e2</guid>
            <category><![CDATA[budgeting]]></category>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[marketing]]></category>
            <dc:creator><![CDATA[Obsidian Studios]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 21:54:42 GMT</pubDate>
            <atom:updated>2026-11-19T21:54:42.259Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QeKwr7AZmeHBZn9b.png" /></figure><p>Quotes for a 30-second AI ad film in India range from a few thousand rupees to several lakh. The difference is rarely the software. It is everything around it.</p><h3>The idea and the script</h3><p>The most valuable part of any ad is the idea. A clear script, written for the platform where the film will run, decides whether the film works before a single frame exists. Cheap quotes often skip this step and start from a prompt.</p><h3>The board</h3><p>A storyboard or animatic lets you approve the film before production. It also tells the studio exactly which shots are needed, which keeps generation focused and the budget under control.</p><h3>Generation</h3><p>Tool subscriptions and compute are a real cost, but usually a small share of the total. What costs more is the number of attempts per shot: a film that holds up on a large screen needs many takes, most of which are thrown away.</p><h3>Edit, sound and colour</h3><p>Finishing turns a set of clips into a film. Editing, sound design, music and a colour grade take the same skill they always have, and they are the steps most often missing from the cheapest quotes.</p><h3>Product accuracy</h3><p>If your product appears, its shape, label and colours must be right in every frame. Checking and correcting that takes time, and it rises with how much the product is on screen.</p><h3>Versions</h3><p>One master film can become a 15-second cut, a 6-second bumper and vertical, square and wide versions. Each version adds some editing time, but far less than making a new film.</p><h3>How to compare quotes</h3><p>Ask every studio the same questions: is the script included, is there a board, how many revision rounds, which versions and formats are delivered, who owns the final film, and which tools are used. Two quotes that answer those the same way can be compared fairly.</p><p>If you would like a quote for your brief, message us on <a href="https://wa.me/917977003486">WhatsApp</a>.</p><p><em>Originally published at </em><a href="https://obsidianstudios.in/intelligence/ai-ad-film-cost-explained"><em>https://obsidianstudios.in</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b36cf93213e2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to keep one character consistent across a whole AI campaign]]></title>
            <link>https://medium.com/@obsidianstudios/how-to-keep-one-character-consistent-across-a-whole-ai-campaign-701bb2c8e3a2?source=rss-621e7399d7a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/701bb2c8e3a2</guid>
            <category><![CDATA[branding]]></category>
            <category><![CDATA[video-production]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Obsidian Studios]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 21:53:17 GMT</pubDate>
            <atom:updated>2026-11-19T21:53:17.359Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*V492OUx5WtbvE3Yn.png" /></figure><p>The first shot of an AI campaign is easy. The thirtieth, with the same face, the same jacket and the same product, is where most campaigns fall apart. Consistency is a process, not a setting.</p><h3>Why characters drift</h3><p>Each generation is a new interpretation. Without a fixed reference, the model rebuilds the face, hair and clothes from the description every time, and small differences add up. By the end of a campaign the brand ambassador looks like three different people.</p><h3>Start with a reference sheet</h3><p>Before any scene is made, we build a reference sheet for each character: the face from the front, three-quarter and side, a full-length pose, and close-ups of anything that must not change, such as a tattoo, a hairstyle or a watch. Every scene is generated from that sheet, not from a written description.</p><h3>Lock wardrobe and props separately</h3><p>Faces get attention, but clothes and props drift too. A jacket gains a pocket, a bottle changes shape. Wardrobe and key props get their own references, and the product is treated as fixed artwork, the same way as on a packshot.</p><h3>Keep a shot list with the references attached</h3><p>Every shot in the plan lists which references it uses. When a scene is regenerated weeks later for a new cut-down, the same references go back in, so the new version matches the old.</p><h3>Review the campaign, not the shot</h3><p>A single shot can pass on its own and still fail beside the others. We review shots side by side in sequence and across formats, because a 9:16 crop can expose a difference that a 16:9 frame hides.</p><h3>When to use a real person instead</h3><p>If a campaign is built around a real person’s face or voice, their written consent comes first, and often a short shoot gives a stronger result than a generated likeness. Consistency is easier when the reference is real.</p><p><em>Originally published at </em><a href="https://obsidianstudios.in/intelligence/consistent-character-ai-campaign"><em>https://obsidianstudios.in</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=701bb2c8e3a2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why most AI ads look AI-made, and how to fix it.]]></title>
            <link>https://medium.com/@obsidianstudios/why-most-ai-ads-look-ai-made-and-how-to-fix-it-f521763ea1cf?source=rss-621e7399d7a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/f521763ea1cf</guid>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[video-production]]></category>
            <dc:creator><![CDATA[Obsidian Studios]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 21:51:47 GMT</pubDate>
            <atom:updated>2026-11-19T21:51:47.092Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*apOKmX0f8JjWbJya.png" /></figure><p>Most AI ads are spotted in the first two seconds. Not because the technology is poor, but because a handful of details were never checked. Here are the seven we look for in every frame.</p><h3>1. Hands</h3><p>Hands are still where generated footage most often breaks: extra fingers, fingers that merge, a grip that does not hold the object. The fix starts in the board. If a shot does not need hands, we frame it so it does not show them. If it does, the action is kept simple and the shot is generated several times until one holds up frame by frame.</p><h3>2. Eyes</h3><p>Viewers read faces before anything else. Eyes that drift, do not converge on what the person is looking at, or never blink make a face feel wrong even when nobody can say why. We check the eyeline against the edit, not just the single shot, because a look that works alone can break the cut next to it.</p><h3>3. Text, labels and logos</h3><p>Generated text warps. On a product film that is fatal: a misspelt label is worse than no film at all. Product names, labels and logos are treated as fixed artwork and checked on every frame they appear in.</p><h3>4. Weight and physics</h3><p>Liquids that pour too slowly, fabric that floats, a bottle that lands without weight. Motion without mass is one of the clearest signals of generated footage. Choosing camera moves and actions the models handle well, then rejecting takes that feel weightless, matters more than any single prompt.</p><h3>5. Light that changes between shots</h3><p>Each shot can look beautiful and still fail as a sequence if the light jumps between them. We set the lighting direction and colour for the whole film before generating, and grade the edit as one piece at the end.</p><h3>6. Skin that is too perfect</h3><p>Real skin has texture, pores and small irregularities. Plastic-smooth skin reads as artificial immediately. We ask for texture rather than perfection, and check close-ups at full resolution, not on a phone preview.</p><h3>7. Sound</h3><p>Silence or generic music is a giveaway of its own. Real films have room tone, footsteps, fabric, the click of a cap. Sound design is often the cheapest way to make a generated film feel shot.</p><h3>The pattern behind all seven</h3><p>None of these are solved by a better model alone. They are solved by direction: a clear board, consistent references, many takes per shot and an editor willing to throw most of them away. That is the part of AI production that still depends entirely on people.</p><p><em>Originally published at </em><a href="https://obsidianstudios.in/intelligence/why-ai-ads-look-ai-made"><em>https://obsidianstudios.in</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f521763ea1cf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your customers already asked an AI about you. Here’s what it told them.]]></title>
            <link>https://medium.com/@obsidianstudios/your-customers-already-asked-an-ai-about-you-heres-what-it-told-them-29577ebbf42b?source=rss-621e7399d7a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/29577ebbf42b</guid>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[seo]]></category>
            <dc:creator><![CDATA[Obsidian Studios]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 21:49:50 GMT</pubDate>
            <atom:updated>2026-11-19T21:49:50.479Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mxr6s6ESRBvb1JyA.png" /></figure><p>A growing share of the people who will buy from you this month have already asked an AI assistant what to buy, before they searched Google or opened your site. If your brand was not in that answer, you were not in the running, and you never found out.</p><h3>The front door moved faster than most brands noticed</h3><p>In February 2026, an Envision Horizons survey of US adults found that 63.1% had used AI for product research, up from 49% a year earlier (reported by eMarketer). ChatGPT itself crossed 900 million weekly users in early 2026, and over the same stretch, publisher referral traffic from Google fell by roughly a third, according to the 5WPR First Stop study. The buyer’s shortlist is now being built before the first click most agencies are still optimising for.</p><h3>Being left out is not neutral</h3><p>When an AI assistant does not know your brand, it does not leave your spot empty. It fills it with whoever it does know. eMarketer’s reporting on the same research found that nearly a third of AI-assisted shoppers use the tool specifically to surface alternatives to a brand they were already considering, which means an invisible brand is not just skipped, it is actively replaced in the conversation a prospective customer is having about them.</p><p>And the two channels are not separate: Semrush’s 2026 buyer-journey study found that when an AI tool does mention a brand, 40% of people then go and search Google to confirm it before deciding. A brand that looks confident and consistent in both places converts that follow-up search; a brand with a thin or contradictory presence loses the customer right there, after having already won the first mention.</p><h3>Why most Indian D2C and brand sites fail this test</h3><p>Three things usually explain it. The site blocks or ignores AI crawlers like GPTBot and OAI-SearchBot instead of allowing them. The business has no structured data telling a machine, plainly, who it is, what it sells and where it operates. And its name, address and socials differ slightly across its website, Google Business Profile, directories and press mentions, so the few independent sources an AI model can cross-check against don’t agree with each other.</p><p>None of these are expensive problems. They are just invisible to a founder who is judging the website by how it looks, not by whether a machine can read it.</p><h3>What we actually check for a brand</h3><p>At Obsidian Studios we run this as a straight audit before recommending anything: can GPTBot and the other major AI crawlers reach the site at all; does the site state clearly, in the first sentence of each page, what the business does; is there Organization, Service and FAQ structured data a model can parse; and do the brand’s name, address and socials match everywhere a model might look for a second opinion. Most of that gets fixed in days. What takes longer, and matters more, is building the few genuine, citable sources — real reviews, real listings, real answer pages — that give an AI model a reason to trust the brand enough to recommend it.</p><h3>Where to start</h3><p>Ask yourself the question your customers are already asking an AI model. Then open ChatGPT or Google’s AI Overview yourself and see what comes back. If your brand isn’t in that answer, or the details are wrong, that is the gap. If you want us to run the audit and tell you plainly what AI tools currently say about you, and what it would take to change it, message us on <a href="https://wa.me/917977003486">WhatsApp</a>.</p><p><em>Originally published at </em><a href="https://obsidianstudios.in/intelligence/ai-is-already-deciding-who-gets-found"><em>https://obsidianstudios.in</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29577ebbf42b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>