<?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 Qingwancong on Medium]]></title>
        <description><![CDATA[Stories by Qingwancong on Medium]]></description>
        <link>https://medium.com/@qingwancong?source=rss-c956225e808e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*8W7wyk8ljH8iPsT8</url>
            <title>Stories by Qingwancong on Medium</title>
            <link>https://medium.com/@qingwancong?source=rss-c956225e808e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 01:28:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@qingwancong/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[The window was the light meter: how to recover a backlit photo without blowing the sky]]></title>
            <link>https://medium.com/@qingwancong/the-window-was-the-light-meter-how-to-recover-a-backlit-photo-without-blowing-the-sky-9175f671d4be?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/9175f671d4be</guid>
            <category><![CDATA[travel-photography]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[portrait-photography]]></category>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[photo-editing]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 10:30:39 GMT</pubDate>
            <atom:updated>2026-11-17T10:30:39.072Z</atom:updated>
            <content:encoded><![CDATA[<h4>A short checklist for lifting faces and foregrounds in photos taken against a bright window or sky, without a halo in the hair.</h4><p>Most backlit photos fail in one move. The camera (or the phone) meters for the bright window, the sky, or the sun behind the person, and the face drops into a dark shape. The picture still has a beautiful background. It just no longer has a readable subject.</p><p>The usual fix is a global exposure or “shadows” slider. That often trades one problem for three: the window turns into a white rectangle, a bright halo appears around the hair, and the darkest parts of the face get a smeared, invented texture that was never in the file.</p><p>A better goal is narrower: lift only the underexposed areas that still hold real detail, and leave the bright background alone.</p><h3>1. Decide what is actually recoverable</h3><p>Open the original and look slowly, not at a thumbnail.</p><p>If you can still see the eyes, the pattern of a shirt, or the edge of a jaw when you brighten a little, that area may be worth lifting.</p><p>If a region is a flat, clipped white — a blown window pane, a sky with no cloud texture — there is nothing honest to recover there. Leave it.</p><p>If a face is almost a silhouette and you cannot make out features at any brightness, do not ask an editor to invent a sharp, confident face.</p><p>Name the parts that should stay bright (the window, the sky, a lamp) and the parts you are trying to read (the face, the foreground). You are not trying to relight the whole room.</p><h3>2. Watch the hairline and other edges</h3><p>Backlit photos fail most often at the edge between a dark subject and a bright background. A heavy shadow lift or a crude “dehalo” pass leaves a glowing outline around hair, shoulders and glasses.</p><p>After any edit, zoom in on those edges. If a bright fringe appeared that was not in the original, the lift was too strong or too global. Pull it back. A slightly darker face with a clean outline is better than a brighter face with a halo.</p><h3>3. Protect the highlights you already have</h3><p>The bright background is usually the only clean, sharp part of a backlit frame. When shadows come up, check that the window muntins, the sky gradient, or the outdoor color did not clip further into flat white.</p><p>If the background was the reason you took the picture, it should still look like the same place, at the same time of day.</p><h3>4. Don’t let recovery invent a person</h3><p>Shadow areas in a backlit phone photo are often thin on information. When an edit “recovers” a face that was mostly dark, it may be guessing — a smoother jaw, a different expression, catchlights that were not there.</p><p>For photos of real people, that guess matters. A useful rule: if you cannot see a feature in the original at any brightness, the edited version should not confidently show it either. Accept a softer, honest face over a crisp, invented one. Do not add objects, and do not change clothing, age or expression.</p><h3>5. Compare side by side, at the size people will see it</h3><p>Put the original and the edit next to each other at phone size and at full screen. Ask:</p><p>Does it still look like the same bright background, not a blown blank?</p><p>Is there a new halo around the hair or shoulders?</p><p>Are the people still recognisably themselves, with the same clothes, age and expression?</p><p>Is the composition untouched — nothing added, nothing removed?</p><p>If any answer is no, pull the edit back. A slightly darker result that passes all four is better than a bright one that fails one.</p><h3>A ready-made starting instruction</h3><p>I keep coming back to the same written instruction for this job, because it puts the limits first:</p><blockquote>Lift only underexposed shadow areas in this backlit scene. Protect highlights and avoid halos. Recover only visible evidence; do not invent detail in fully clipped regions. Preserve every person’s identity, expression, age, clothing and the original composition. Do not invent missing facial detail or add objects.</blockquote><p>That exact instruction is set up as a template on MendTrue, an AI photo enhancement tool, where you can upload a backlit portrait or indoor window shot and adjust the wording before generating: <a href="https://mendtrue.com/photo-templates/backlit-recovery">https://mendtrue.com/photo-templates/backlit-recovery</a></p><p>Whatever tool you use, the checklist is the same: lift only what still holds detail, leave the bright background alone, watch the hairline, and never let “brighter” become “made up”.</p><p><em>Disclosure: I work on MendTrue. No example images are included here; the instruction above is the template’s actual text.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9175f671d4be" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A night photo is supposed to be dark: how to lift the shadows without turning midnight into noon]]></title>
            <link>https://medium.com/@qingwancong/a-night-photo-is-supposed-to-be-dark-how-to-lift-the-shadows-without-turning-midnight-into-noon-11287cb83db6?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/11287cb83db6</guid>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[photo-editing]]></category>
            <category><![CDATA[night-photography]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 10:31:09 GMT</pubDate>
            <atom:updated>2026-11-14T10:31:09.609Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A short checklist for brightening city and landscape night shots while keeping the sky, the signs and the mood.</em></p><p>Most night photos fail in the same quiet way. The phone exposed for the bright shop sign, the streetlight or the moon, and everything you actually cared about — the friends on the bridge, the shape of the street, the trees at the edge of the lake — sank into a muddy dark.</p><p>The instinct is to drag an exposure slider to the right. That usually trades one problem for three: the sky turns grey-brown, the sodium lamps go orange-white, noise blooms in every shadow, and the picture stops looking like night at all. It looks like a badly lit afternoon.</p><p>A better goal is narrower: open up the dark areas that still hold real detail, and leave the night alone.</p><h4>1. Decide what “dark” is allowed to stay</h4><p>Before editing, look at the photo and name three things that should remain dark:</p><ul><li>the sky (it should still read as night sky, not overcast day)</li><li>the deepest corners with no information in them</li><li>the shadow side of buildings or hills, which gives the scene depth</li></ul><p>Then name what needs to be readable: faces, the path, the foreground. You are lifting those, not the whole frame.</p><h4>2. Treat artificial light as several colors, not one cast</h4><p>A single night street can mix warm sodium lamps, cool LED signs and a neutral shop window. A global “fix white balance” pass picks one of them as “correct” and pushes the others into strange colors.</p><p>Aim for balance instead of neutrality: tame the strongest cast a little so skin doesn’t look orange, and keep the neon looking like neon. If every light source ends up the same white, the photo has lost its location.</p><h4>3. Protect the highlights you already have</h4><p>Illuminated signs, windows and lamps are usually the only clean, sharp parts of a night photo. When you brighten shadows, check that these don’t clip into flat white blobs or lose their lettering. If a sign was readable in the original, it should still be readable after the edit.</p><h4>4. Don’t let brightening invent detail</h4><p>Shadow areas in phone night shots are often more noise than information. When an edit “recovers” a face that was mostly black, it may be guessing — a smoother jaw, a different expression, eyes that weren’t really visible.</p><p>For photos of real people, that guess matters. A useful rule: if you can’t see a feature in the original at any brightness, the edited version shouldn’t confidently show it either. Accept a softer, honest face over a crisp, invented one. Same for objects: no new lamps, no extra stars, no cleaned-up crowd.</p><h4>5. Compare side by side, at the size people will see it</h4><p>Put the original and the edit next to each other at phone size and at full screen. Ask:</p><ul><li>Does it still look like the same time of night?</li><li>Is the sky still a night sky?</li><li>Are the people still recognisably themselves, with the same clothes, age and expression?</li><li>Is the composition untouched — nothing added, nothing removed?</li></ul><p>If any answer is no, pull the edit back. A slightly darker result that passes all four is better than a bright one that fails one.</p><h4>A ready-made starting instruction</h4><p>I keep coming back to the same written instruction for this job, because it puts the limits first:</p><blockquote>Gently lift dark areas and balance artificial light colors in this night scene. Preserve the dark sky, illuminated signs and nighttime atmosphere; do not turn it into daylight. Preserve every person’s identity, expression, age, clothing and the original composition. Do not invent missing facial detail or add objects.</blockquote><p>That exact instruction is set up as a template on MendTrue, an AI photo enhancement tool, where you can upload a city or landscape night photo and adjust the wording before generating: <a href="https://mendtrue.com/photo-templates/night-scene-balance">https://mendtrue.com/photo-templates/night-scene-balance</a></p><p>Whatever tool you use, the checklist is the same: keep the night, lift only what holds real detail, balance the lights instead of erasing them, and never let “brighter” become “made up”.</p><p><em>Disclosure: I work on MendTrue. No example images are included here; the instruction above is the template’s actual text.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=11287cb83db6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Keep the Crease Story Out of the Face: A Checklist for Gentle Photo Restoration]]></title>
            <link>https://medium.com/@qingwancong/keep-the-crease-story-out-of-the-face-a-checklist-for-gentle-photo-restoration-4cd36fad2353?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4cd36fad2353</guid>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[digital-archives]]></category>
            <category><![CDATA[photo-editing]]></category>
            <category><![CDATA[family-history]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 11:40:25 GMT</pubDate>
            <atom:updated>2026-11-04T11:40:25.380Z</atom:updated>
            <content:encoded><![CDATA[<h4>Repair dust, scratches, and fading without smoothing away the person who was actually there.</h4><p>Family albums hold photographs that cannot be retaken. A crease cuts across a cheek. White flecks sit in the hairline. The edges are scuffed from decades in a shoe box. Someone wants a cleaner shareable copy, and an automatic “restore” button looks like the fastest path.</p><p>The risk is not that the tool fails to remove dust. The risk is that it also removes the person: smoother skin, a younger smile, a sharper collar that never existed, a backdrop rebuilt like a stock portrait. If the point of the print is who that person was on that day, those changes matter more than a spotless scan.</p><p>This piece is a practical checklist for gentle restoration: fix visible age damage, keep original grain and tone, and verify identity before you share or print.</p><h3>What “gentle” actually means</h3><p>A conservative restore does three things and refuses a fourth.</p><p>It removes or softens damage that is clearly on the surface of the print or scan: dust, small scratches, fold lines, mild fading.</p><p>It keeps the existing color or monochrome treatment. A black-and-white studio portrait should stay black and white unless you deliberately open a separate colorization pass.</p><p>It preserves identity, expression, age, clothing, and composition.</p><p>It does not invent missing facial detail or add objects. If a crease erased part of an ear, the honest answer may be a soft repair — not a newly drawn ear that looks like a different person.</p><p>When a tool advertises “AI restore,” check which of those four it actually does. Many packages optimize for a flattering modern portrait. That is a different job from archival cleanup.</p><h3>Repair damage before you ask for beauty</h3><p>Separate the jobs. First pass: dust, scratches, crease, mild fade. Second pass, only if you still need it: tone, crop, or a labelled color preview on a duplicate file.</p><p>Running beauty, colorization, and damage repair in one click makes failures hard to diagnose. When the face drifts, you cannot tell whether the model over-smoothed skin or reinvented the mouth while filling a crease. Keep a repaired master that still looks like the same photograph, then decide whether any further edit is worth the risk.</p><h3>A worked example</h3><p>Here is the demo pair from MendTrue’s gentle restoration template. Both images are official MendTrue demo assets and are AI-generated sample images used on the template page, not a real person and not a customer photo. The input is a mid-century style studio headshot with visible age damage.</p><figure><img alt="A damaged black-and-white mid-century studio portrait of a young woman with waved dark hair and a light pointed collar; a hard horizontal crease crosses the cheeks and ears; dust flecks and edge scuffs are visible" src="https://cdn-images-1.medium.com/max/1024/1*ZRWWOdi5fGJaydQL0w2V3A.jpeg" /><figcaption>INPUT: official MendTrue demo image (photo-restoration-before) from the gentle restoration template. AI-generated sample, not a real person.</figcaption></figure><p>The input still has clear structure: soft directional studio light, a mottled grey backdrop, dark mid-century waves, a light pointed collar, a calm expression. Across the middle sits a hard horizontal crease through the cheeks and ears. Fine white flecks and scuffs sit on the surface. That crease and those flecks are the restore target — not the hairline, not the collar shape, not the age of the face.</p><figure><img alt="The same studio portrait after a conservative restore — crease and surface flecks removed, while pose, hairline, collar and expression stay the same; still black and white" src="https://cdn-images-1.medium.com/max/1024/1*0K0JndY1WS5R4Vk3DvqSuA.jpeg" /><figcaption>RESULT: official MendTrue demo output (photo-restoration-after) for the same template. AI-generated example image.</figcaption></figure><p>In the result, the crease and the surface flecks are gone, while the parting, the wave above the forehead, the ear contour, the collar points, and the expression match the damaged input. The portrait stays black and white. Grain and the soft studio character remain. That is the behaviour to look for: damage lifted off the photograph, identity left alone.</p><p>The template’s prepared instruction says this outright: repair visible dust, small scratches and fading conservatively; keep original grain and the existing color or monochrome treatment; preserve every person’s identity, expression, age, clothing and the original composition; do not invent missing facial detail or add objects. A demo chosen for a template page is also a friendly case. Your own scans will be messier, so check them.</p><h3>How to check a restoration</h3><p>You do not need specialist software. You need the original next to the result and a few minutes.</p><ol><li>Flip between before and after at the same size. Damage should recede without the face sliding into a different person.</li><li>Check hairline, eye shape, teeth if visible, ear contours, and the edge of glasses. Identity drift shows up there first.</li><li>Check skin and fabric texture. Plastic-smooth skin or a rebuilt blouse usually means the model painted a new surface rather than cleaning the old one.</li><li>Check whether monochrome stayed monochrome, or whether color quietly appeared. Accidental colorization is a different edit and should be labelled as such.</li><li>Ask whether you would still recognise the person from the damaged scan alone. If recognition now depends on the “restored” face, the edit went too far.</li></ol><h3>When to keep the damaged master</h3><p>Sometimes the honest share is a lightly cleaned scan, or no automated restore at all. If the print is the only copy of a relative, if a legal or archival use needs the unaltered record, or if every automatic pass changes the face, keep the original scan as master. Save any restored version as a labelled derivative. Creases and dust tell part of the object’s history; the face should not be rewritten to erase them.</p><h3>Try a conservative pass on your own scan</h3><p>If you have a print with dust, small scratches, or mild fading that you want cleaned without rewriting the person, the <a href="https://mendtrue.com/photo-templates/gentle-restoration">gentle restoration template on MendTrue</a> is set up for this job, with the cautious instruction above already filled in and editable. Whatever tool you use, keep the original scan, save the restored version as a labelled copy, and run the checks above before you share or print it.</p><p>Gentle restoration is cleanup, not reinvention. The goal is a photograph you can still trust when you put it next to the original.</p><p>Disclosure: both images in this article are official MendTrue demo assets and are AI-generated example images, not a real person and not a customer photo. I work on MendTrue.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4cd36fad2353" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Check CUDA Before You Burn the Night: A First-Run Checklist for MicroDuck RL]]></title>
            <link>https://medium.com/@qingwancong/check-cuda-before-you-burn-the-night-a-first-run-checklist-for-microduck-rl-716bb9a6dcd7?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/716bb9a6dcd7</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[open-source]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 11:30:18 GMT</pubDate>
            <atom:updated>2026-11-01T11:30:18.482Z</atom:updated>
            <content:encoded><![CDATA[<h4>An unofficial walkthrough of the source-reviewed local-training setup — pin a revision, verify the GPU stack, then run a short install check.</h4><p>Hobby robot training nights fail in boring ways. The wrong Python lands in the path. The lockfile silently drifts. CUDA looks installed until the first import. By the time you notice, you have already burned an evening and a cloud budget on a setup that never finished the first gate.</p><p>This is a first-run checklist for the official MicroDuck RL CUDA workflow, written for builders who want to confirm the environment before a full training session. It is based on the source-reviewed guide on MicroDuck Robot Community, an independent unofficial site. It is not affiliated with Pollen Robotics or Hugging Face, and it does not claim to have run a full GPU training benchmark.</p><h3>Pin one reviewed revision in its own checkout</h3><p>Do not train against whatever main happens to be tonight. Clone into a fresh directory, detach at a reviewed commit, and sync with a locked dependency set under Python 3.12.</p><p>The walkthrough pins revision 2b25a48 and uses uv sync --locked. Locked sync refuses to quietly resolve a new dependency graph when the lockfile and project disagree. If sync fails, keep the error instead of deleting the lockfile and hoping.</p><p>Python 3.13 is out of scope for this reviewed route. A detached checkout also keeps the session on one revision so an accidental pull does not rewrite the night’s baseline.</p><h3>Ask the environment the questions that matter</h3><p>Before you schedule a long job, print what the training process will actually see: Python version, machine arch, PyTorch build, whether CUDA is available, device count, and the first GPU name if any. For this CUDA path you want Python 3.12.x, a CUDA build of PyTorch, CUDA available: True, and at least one visible GPU.</p><p>That check proves discovery. It does not prove memory headroom or end-to-end training compatibility. If importing PyTorch itself fails, keep the traceback and fix dependencies before you invent a new command line.</p><p>Then list the installed training tasks and confirm the expected MicroDuck velocity task name is present. Missing task names usually mean the wrong checkout or an incomplete install, not a typo worth papering over.</p><h3>Prefer a short install check over a heroic first run</h3><p>A brief installation check — a few steps on the registered task, with logging you can read — is cheaper than discovering a broken stack after hours of silent compute. Treat that short run as an editorial smoke test of the install, not as an upstream performance claim.</p><p>If the short run fails, you still have a small log and a pinned revision. If it succeeds, you earned the right to commit a longer session on the same checkout.</p><h3>Where the full commands live</h3><p>The full command blocks, source links, and troubleshooting notes are on the <a href="https://microduckrobot.org/guides/local-training/">MicroDuck local training guide.</a> Use that page as the single source for copy-paste steps. This article is the mindset and the order of checks; the guide is the exact sequence.</p><h3>What this checklist does not claim</h3><p>It does not claim shipping timelines, user counts, or benchmark scores. It does not speak for the upstream project. MicroDuck Robot Community reviews public sources and writes independent notes; GPU training on this site’s own hardware is not claimed here.</p><p>If you are about to start a local MicroDuck RL night, pin the revision, verify CUDA from that environment, confirm the task registry, and run the short install check first. The long job can wait until those four lights are green.</p><p>Disclosure: MicroDuck Robot Community (microduckrobot.org) is an independent unofficial community site. It is not affiliated with Pollen Robotics, Hugging Face, or the upstream MicroDuck maintainers. I help operate that community site. The revision pin and CUDA workflow above follow the site’s source-reviewed guide; they are not an official endorsement and not a performance result from a full training run.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=716bb9a6dcd7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Color Is a Guess: How to Preview a Monochrome Print Without Rewriting the Face]]></title>
            <link>https://medium.com/@qingwancong/color-is-a-guess-how-to-preview-a-monochrome-print-without-rewriting-the-face-2522be16aae7?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/2522be16aae7</guid>
            <category><![CDATA[photo-editing]]></category>
            <category><![CDATA[family-history]]></category>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[digital-archives]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 10:50:08 GMT</pubDate>
            <atom:updated>2026-10-29T10:50:08.489Z</atom:updated>
            <content:encoded><![CDATA[<h4>Treat colorization as an interpretation of luminance, not a claim about what the original dyes looked like.</h4><p>Old family albums are full of black-and-white prints that someone, sooner or later, wants to “see in color.” A sibling asks for a warmer shareable version. A cousin wants to print a birthday poster. A repair shop offers one-click colorization as part of a restore package.</p><p>The risk is not that color looks wrong in a taste sense. The risk is that the tool quietly rewrites the photograph: smoother skin, a younger smile, a different hairline, clothing that never existed. If the point of the print is who that person was on that day, those changes matter more than a nicer palette.</p><p>This piece is about restraining colorization: keep the monochrome structure, add only plausible color, and check the result the way you would check a restoration.</p><h3>Colorization fills in what the print never recorded</h3><p>A black-and-white photograph records luminance. It does not record hue. Any color you add later is an interpretation built on top of brightness and shape. That is fine when you treat it as a preview. It becomes a problem when the result is presented as historical fact, or when the model invents detail the scan never held.</p><p>Two failure modes show up again and again.</p><p>Over-painting. The tool does not only tint. It rebuilds faces and fabrics so they look like a modern portrait. Identity drifts even when the overall pose stays the same.</p><p>Period costume. Hair, lipstick, and clothing get the saturated look of a stock illustration rather than the muted dyes of a mid-century print. The photo starts to look like a different decade.</p><p>A restrained edit does the opposite. It keeps period detail and luminance stable, adds modest color where the structure already supports it, and refuses to invent missing facial information.</p><h3>Repair first, color second</h3><p>If the scan is dusty, scratched, or soft, fix that on a separate pass before you ask for color. Colorization on top of damage often paints over the damage instead of leaving you a clean base you can still trust. Keep the repaired monochrome master. Save the color version as a labelled copy, not as a replacement of the original scan.</p><h3>A worked example</h3><p>Here is the demo pair from MendTrue’s restrained colorization template. Both images are AI-generated sample images used on the template page, not a real person and not a customer photo. The input is a mid-century style studio headshot in black and white.</p><figure><img alt="A black-and-white studio portrait of a young woman with waved dark hair, a light collared shirt and pearl stud earrings." src="https://cdn-images-1.medium.com/max/1024/1*eX_kEHAisB6jO4o_i8C29g.jpeg" /><figcaption><em>INPUT: official MendTrue demo image (colorize-before) from the restrained colorization template. AI-generated sample, not a real person.</em></figcaption></figure><p>The input has clear structure to work from: a soft studio backdrop, directional light on the face, polished mid-century waves, a light collared shirt, and small pearl earrings. There is</p><p>no color information to recover. Anything that appears in color in the next frame is an interpretation.</p><figure><img alt="The same studio portrait with restrained color — muted blue shirt, warm skin, dark brown hair, soft lipstick — while pose, hairline and expression stay the same." src="https://cdn-images-1.medium.com/max/1024/1*KX6FMiNrUTKQXcOMVzjAlg.jpeg" /><figcaption><em>RESULT: official MendTrue demo output (colorize-after) for the same template. AI-generated example image.</em></figcaption></figure><p>In the result, the shirt reads as a muted blue, the hair as dark brown, and the lips as a soft warm red, while the backdrop stays a quiet grey-brown. Just as important is what did not change. The parting, the wave above the forehead, the ear shape, the collar points and the expression match the monochrome input. The edit added color on top of the existing luminance instead of redrawing the person.</p><p>That is the behaviour to look for. The template’s prepared instruction says this outright: repair minor dust first, then add restrained, plausible color; keep period details and luminance stable; treat color as an interpretation, not a claim about historical accuracy; preserve identity, expression, age and clothing; do not invent missing facial detail or add objects. A demo chosen for a template page is also a friendly case. Your own scans will be messier, so check them.</p><h3>How to check a colorization</h3><p>You do not need specialist software. You need the original next to the result and a few minutes.</p><ol><li>Flip between before and after at the same size. Color should appear without the face sliding into a different person.</li><li>Check hairline, eye shape, teeth if visible, ear contours and the edge of glasses. Those are where identity drift shows up first.</li><li>Check clothing texture. A flat, plastic-looking fabric usually means the model painted a new surface rather than tinting the old one.</li><li>Check saturation. Mid-century prints rarely look like modern phone portraits. If every hue is vivid and even, dial the edit back or reject it.</li><li>Ask whether you would still recognise the person from the monochrome alone. If the answer depends on the new color, the edit went too far.</li></ol><h3>When to keep the archive black and white</h3><p>Sometimes the honest answer is not to colorize at all. If the scan is the only copy of a relative, if a legal or archival use needs the unaltered record, or if every automatic color pass changes the face, keep a restored monochrome master and share that. Color can wait for a labelled experiment on a duplicate file.</p><h3>Try a restrained preview on your own scan</h3><p>If you have a monochrome print you want to preview in color without rewriting the person, the <a href="https://mendtrue.com/photo-templates/restrained-colorization">restrained colorization template on MendTrue</a> is set up for this job, with the cautious instruction above already filled in and editable. Whatever tool you use, keep the original scan, save the color version as a labelled copy, and run the checks above before you share or print it.</p><p>Color is a guess laid on top of light. The goal is a preview you can distrust productively, not a new face with nicer hues.</p><p>Disclosure: both images in this article are official MendTrue demo assets and are AI-generated example images, not a real person and not a customer photo. I work on MendTrue.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2522be16aae7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Glare Is Not Damage: How to Get a Clean Copy of a Framed Photo With Your Phone]]></title>
            <link>https://medium.com/@qingwancong/glare-is-not-damage-how-to-get-a-clean-copy-of-a-framed-photo-with-your-phone-838078d0dc0b?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/838078d0dc0b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[smartphone-photography]]></category>
            <category><![CDATA[family-history]]></category>
            <category><![CDATA[photo-editing]]></category>
            <category><![CDATA[photography]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 10:36:09 GMT</pubDate>
            <atom:updated>2026-10-26T10:36:09.842Z</atom:updated>
            <content:encoded><![CDATA[<h4>A reflection sits on top of the print, not inside it. Prevent what you can, clean up what survives, and re-shoot what doesn&#39;t.</h4><p>Many family photos that matter are still prints: in frames on a shelf, behind glass, in albums with plastic sleeves. When you want to send one to a sibling or keep a digital copy, the fastest route is your phone camera. And the most common thing that goes wrong is not blur or color. It is glare.</p><p>A window reflected across the glass. A ceiling light turned into a bright dot right next to someone&#39;s face. A haze of grey over one corner where the frame caught the room. The print itself is fine. The copy you just made is not.</p><p>This piece is about getting a usable copy anyway: what you can prevent at capture time, what software can reasonably clean up afterwards, and where the honest answer is &quot;take the picture again.&quot;</p><h3>Glare is light on top of the photo, not damage in it</h3><p>It helps to be precise about what glare is. A reflection does not change the print underneath. It adds light on top of it. In practice that light shows up in two different ways.</p><p>Partial glare is a veil. You can still see the picture through it, just washed out and lower in contrast. A reflected window frame often looks like this: a pale diagonal band with the original image still visible underneath.</p><p>Full glare is a blowout. The reflection is so bright that the camera recorded pure white, and nothing of the print survives in those pixels. Small, hard hotspots from lamps and ceiling lights are often like this in their centre.</p><p>That distinction decides almost everything. Where detail survives under the veil, it can be recovered. Where the pixels are white, any &quot;recovered&quot; detail is a guess.</p><h3>First, try not to need a fix</h3><p>Five minutes at capture time beats any amount of editing.</p><ol><li>Take the photo out of the frame if you can. Glass is the main source of reflections, and most prints are matte or semi-matte on their own.</li><li>Turn off overhead lights and shoot near a window with indirect daylight, with the window beside you, not behind you.</li><li>Tilt the phone slightly, by 10 to 15 degrees, rather than shooting perfectly straight on. Reflections move much more than the print does. Correct the small perspective change later with a crop.</li><li>Watch the screen, not the print. Move until the bright patch slides off the faces and important areas, then shoot.</li><li>Take three or four versions from slightly different angles. Different shots carry glare in different places, which gives you options.</li></ol><p>Sometimes you only get one quick shot of a print on someone else&#39;s wall. That is where cleanup comes in.</p><h3>A worked example</h3><p>Here is the demo pair from MendTrue&#39;s printed photo glare template. Both images are AI-generated sample images used on the template page, not a real person and not a customer photo. The input is a framed portrait, photographed on a shelf, with the room reflected in the glass.</p><figure><img alt="A framed portrait photographed on a shelf, with a window reflection, three round light hotspots and grey haze on the glass" src="https://cdn-images-1.medium.com/max/1024/1*nLUopQSAWBMx3OYORZBtLw.jpeg" /><figcaption>INPUT: official MendTrue demo image (photo-glare-before) from the printed photo glare template. AI-generated sample, not a real person.</figcaption></figure><p>The glare in the input is a fairly typical mix: a pale reflection of a window frame across the upper left of the photo, a grey haze over the plant in the lower left, and three round hotspots from room lights, two of them close to the person&#39;s head.</p><figure><img alt="The same framed portrait with the window reflection, hotspots and haze removed; face, clothing, frame and mat unchanged" src="https://cdn-images-1.medium.com/max/1024/1*15CmFf2Qr73IOgymkCN-YQ.jpeg" /><figcaption>RESULT: official MendTrue demo output (photo-glare-after) for the same template. AI-generated example image.</figcaption></figure><p>In the result, the window reflection and the three hotspots are gone, and the background reads again as a plain warm wall with a plant beside it. Just as important is what did not change. The face, the smile, the hairline, the green top and its neckline are the same. The frame, the white mat and the shelf are untouched. The edit stayed on the reflection and left the photograph alone.</p><p>That is the behaviour to look for. The template&#39;s prepared instruction says this outright: reduce surface glare only where the underlying detail is still visible, keep the image content unchanged, and do not reconstruct obscured faces or invent missing information. A demo picked for a template page is also a friendly case. Your own photos will be messier, so check them.</p><h3>How to check a glare cleanup</h3><p>You do not need any special skill to judge the result. You need a few minutes and the original next to it.</p><ol><li>View the before and after at the same size and flip between them. Changes are much easier to see when you toggle than when you compare side by side.</li><li>Look at every place the glare used to be. Does the area now match its surroundings in texture and tone, or is there a smooth, flat, slightly blurry patch? A flat patch usually means the tool painted over rather than recovered.</li><li>Check faces especially carefully if a hotspot was on or near one. If an eye, a tooth, or the edge of glasses was inside a white hotspot, whatever you see there now was generated. Decide whether you can live with that.</li><li>Check everything the glare never touched. Faces, clothing and backgrounds away from the reflection should be essentially unchanged. If an untouched face suddenly looks smoother, younger or sharper, the tool is rewriting the photo, not cleaning it.</li><li>Compare the overall tone. Glare lowers contrast, so some gain in contrast is expected where the veil was removed. A whole-image jump in saturation or sharpness is not part of glare removal.</li></ol><h3>When to stop and re-shoot</h3><p>Cleanup works best on partial glare over backgrounds, clothing and other low-stakes areas. It is far less trustworthy when a solid white reflection covers the centre of a face or a line of handwriting on the print. In those cases, the better answer is almost always another photo from a different angle, even if you have to ask a relative to take the print out of the frame for you.</p><h3>Try it on your own copy</h3><p>If you have a phone photo of a framed print with a reflection across it, the <a href="https://mendtrue.com/photo-templates/photo-print-glare">printed photo glare template on MendTrue</a> is set up for this specific job, with the cautious instruction above already filled in and editable. Whatever tool you use, keep your original phone shot, save the cleaned version as a labelled copy, and run the checks above before you share or print it.</p><p>A reflection is the room, not the photo. The goal is to take the room back out without changing the photograph.</p><p>Disclosure: both images in this article are official MendTrue demo assets and are AI-generated example images, not a real person and not a customer photo. I work on MendTrue.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=838078d0dc0b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Torn Photo Is Mostly Still There: How to Repair Tears Without Inventing the Past]]></title>
            <link>https://medium.com/@qingwancong/a-torn-photo-is-mostly-still-there-how-to-repair-tears-without-inventing-the-past-e10f98dbcb1c?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e10f98dbcb1c</guid>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[family-history]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 10:35:51 GMT</pubDate>
            <atom:updated>2026-10-23T10:35:51.939Z</atom:updated>
            <content:encoded><![CDATA[<h4>Most tears are seams, not holes. Close the line, keep what survived, and check the result before you print it.</h4><p>A tear in an old print is one of the most common kinds of damage in a family archive, and one of the most misunderstood. People tend to treat it as a hole that needs filling. Usually it is something smaller and more honest than that: a line where the paper split, with almost everything on both sides still present.</p><p>What a tear actually is</p><p>When a print tears, three things usually happen at once.</p><p>First, the paper splits along a line. The image on each side of that line is mostly intact; it has just been pulled a fraction of a millimetre apart, or folded so the two edges no longer meet cleanly.</p><p>Second, the emulsion along the edge lifts or flakes. That is why a tear on a scan looks like a bright white crack rather than a thin dark line. You are seeing the paper base, not missing picture.</p><p>Third, sometimes, a small piece goes missing. A corner chips off, or a sliver along the tear is gone. This is the only part of the damage where image information is genuinely absent.</p><p>Most tears are overwhelmingly the first two kinds. The picture is still there. The repair job is mostly about closing a seam, not inventing what was underneath it.</p><p>Why you should not invent missing detail</p><p>Modern AI tools are very good at producing plausible pixels. That is exactly the risk. If a tear runs through a background wall, a curtain or a patterned dress, filling it convincingly is low stakes: the texture continues, and nobody&#39;s memory depends on the exact fold of a curtain.</p><p>If a tear runs through a face, an inscription, a hand, or anything that identifies a person or a place, &quot;plausible&quot; is not good enough. A generated eyebrow or an invented letter in a shop sign can look perfectly natural and still be wrong. Once that version gets printed, framed and shared with relatives, the invented detail becomes the record.</p><p>So a good rule is: rebuild only what the surrounding image clearly supports. Continue the background. Close the seam. Match the tone across the split. Where something is truly missing and important, leave it soft, or mark it, rather than giving it a confident new shape.</p><p>A worked example</p><p>Here is the official demo pair from the small tear repair template. The input is a sepia studio portrait with a long tear running from the top edge down through the background and the dress, a few finer cracks across the curtain, and a chipped area near the bottom right.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LbS1_u3bqhQspU6jb428Ig.jpeg" /><figcaption>INPUT: official MendTrue demo image (torn-before), an AI-generated sample portrait with a simulated tear, not a real family photo.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JMK7cXrZ93SuV0Ac72szEg.jpeg" /><figcaption>RESULT: official MendTrue demo output (torn-after) from the small tear repair template. AI-generated example image.</figcaption></figure><p>A few things are worth noticing in the result.</p><p>The long seam across the left background and down through the dress is closed, and the floral pattern continues across where the white crack used to be. The fine cracks in the curtain on the right are gone, and the curtain folds keep their original direction. The face, which the tear never crossed, is left alone: same expression, same features, same earrings. The overall sepia tone and grain stay close to the original rather than being &quot;upgraded&quot; into a glossy modern look.</p><p>How to check a repaired result</p><p>You do not need a trained eye to judge a tear repair.</p><p>Put them side by side at the same size. Flip between them rather than viewing them separately. Changes are easier to see in motion.</p><p>Follow the tear line end to end. Along the whole path, does the texture on one side continue naturally into the other? Look for smears, repeated patches or a soft &quot;blurred stripe&quot; where the crack was.</p><p>Check everything the tear did not touch. Faces, hands, jewellery, text and backgrounds away from the damage should be essentially unchanged. If a face that was never damaged now looks smoother or younger, the tool has been rewriting, not repairing.</p><p>Look hard at anything that was genuinely missing. Chipped corners and lost slivers are where invention happens. Ask whether you could defend that filled area if a relative asked &quot;was that really there?&quot;</p><p>Compare tone and grain. A good repair keeps the original&#39;s colour cast and texture. A sudden jump in sharpness or a cleaner, flatter tone usually means generated content.</p><p>Keep the original scan. Always. Save the repaired version as a copy, and label it as repaired so nobody mistakes it for the untouched print later.</p><p>When to stop</p><p>Small tears through backgrounds and clothing are the easy, low-risk case, and that is where automated repair is most useful. Large missing areas through a face are different. There, the honest outcome may be a partial repair with the gap softened, or a note on the back of the print saying what was lost, rather than a perfect-looking image built on a guess.</p><p>If you want to try this on your own scans, <a href="https://mendtrue.com/photo-templates/small-tear-repair">the small tear repair template on MendTrue</a> is set up for exactly this kind of job: aligning the visible fragments, reducing seam lines and rebuilding small gaps while staying cautious about detail that is truly missing. Whatever tool you use, run the checks above before you print anything.</p><p>A torn photograph is still that photograph. The goal of repair is to make it easier to look at, not to replace it with a more convenient version.</p><p>Disclosure: both images are official MendTrue demo assets and are AI-generated examples, not a real customer photo.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e10f98dbcb1c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Age marks on a print are surface damage — not a reason to invent a younger face]]></title>
            <link>https://medium.com/@qingwancong/age-marks-on-a-print-are-surface-damage-not-a-reason-to-invent-a-younger-face-4961cf266ac5?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4961cf266ac5</guid>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[history]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 10:51:44 GMT</pubDate>
            <atom:updated>2026-10-18T10:51:44.460Z</atom:updated>
            <content:encoded><![CDATA[<p>Age marks on a print are surface damage — not a reason to invent a younger face</p><p>An old print often arrives with the wrong ask attached: “make them look young again,” “fix the face,” “restore the original beauty.” Those prompts treat age damage as an identity problem. Most of the time it is a surface problem.</p><p>Creases, dust, mild yellowing, and soft print wear sit on the paper layer. The person underneath is usually still readable. The dangerous edit is the one that rebuilds cheeks, eyes, and skin tone into a younger stranger while claiming archival recovery.</p><p>I did not use a private family photo. For this check I used MendTrue’s public synthetic archive sample from the site (the same pair the gentle-restoration template shows as its example). It is a demo file, not a real household album.</p><figure><img alt="Official MendTrue synthetic archive sample used as gentle-restoration example input: aged print with visible surface wear" src="https://cdn-images-1.medium.com/max/1024/1*ZRWWOdi5fGJaydQL0w2V3A.jpeg" /></figure><p><em>Input: official MendTrue synthetic archive sample used by the gentle-restoration template. Studio/demo image, not a private family photograph.</em></p><p>The brief was narrow: reduce only the visible surface damage that still leaves the underlying structure intact, keep age cues and expression, and refuse to invent missing facial geometry or a “studio remake” look.</p><figure><img alt="AI gentle-restoration result for the same synthetic sample: reduced visible age damage while keeping original character" src="https://cdn-images-1.medium.com/max/1024/1*0K0JndY1WS5R4Vk3DvqSuA.jpeg" /></figure><p><em>Result from the current gentle-restoration path. This is an AI surface-repair estimate, not a second sitting with a younger subject.</em></p><h3>I start by separating paper from person</h3><p>Before I touch “beauty,” I ask what layer is broken.</p><p>If hairline, glasses, clothing seams, and background geometry are still present under the wear, the job is restrained cleanup of the print surface. If a cheek is a pure blank tear with no residual structure, inventing skin there is fiction — useful sometimes as illustration, never as archival evidence.</p><p>For this sample the person was still readable. So the test is: can the tool calm the age marks without rewriting who is in the frame.</p><h3>Keep character harder than you “fix” damage</h3><p>The original character is the constraint. Once skin becomes plastic, teeth brighten into a new smile, or soft-focus beauty lighting appears, the edit stopped being restoration and started being a portrait remake.</p><p>In review I zoomed on:</p><ul><li>forehead and cheek texture — calmer marks, not a new face;</li><li>eye shape and gaze — same person, not a remade portrait;</li><li>clothing edges and background objects — geometry unchanged;</li><li>color cast — softened only where print aging explained it, not a fashion grade;</li><li>overall age cues — still the same generation of sitter.</li></ul><p>If any of those move into a prettier stranger, I treat the result as failed even if the print “looks cleaner.”</p><h3>What I refuse to ask in the same pass</h3><p>I do not stack “make them younger,” “colorize the whole album,” “beautify skin,” or “cinematic remaster” on top of gentle restoration. Each of those is a different product decision and a different credit.</p><p>I also do not ask for a studio remake. A remake invents a sitting the original camera never had. The useful question is smaller: calm the paper damage that still leaves evidence, and leave missing regions alone.</p><p>You can repeat the same narrow check on MendTrue’s <a href="https://mendtrue.com/photo-templates/gentle-restoration">gentle photo restoration template</a>: one synthetic sample, one restrained repair instruction, and a hard rule not to invent a younger face.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4961cf266ac5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A bright sky and a flat face is an exposure problem — not a reason to invent a new person]]></title>
            <link>https://medium.com/@qingwancong/a-bright-sky-and-a-flat-face-is-an-exposure-problem-not-a-reason-to-invent-a-new-person-0d08b5f89f64?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/0d08b5f89f64</guid>
            <category><![CDATA[travel]]></category>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 10:39:27 GMT</pubDate>
            <atom:updated>2026-10-15T10:39:27.519Z</atom:updated>
            <content:encoded><![CDATA[<p>A coastal selfie against a bright sky often looks “wrong” in the worst way: the water and clouds still have detail, but the face and path sit in a flat, hazy shade. The reflex is to ask AI to “fix the person” or “make it look like golden hour.”</p><p>That reflex is how you get a second, invented face. Backlight is usually recoverable shadow plus protected highlights — not a missing identity problem.</p><p>I did not use a private family photo. For this check I used MendTrue’s public synthetic travel-selfie sample from the site (the same pair the backlit-recovery template shows as its example). It is a demo file, not a real trip archive.</p><figure><img alt="Official MendTrue synthetic travel-selfie sample used as backlit-recovery example input: coastal selfie against bright sky, face and path underexposed" src="https://cdn-images-1.medium.com/max/1024/1*HIWKPjTYSWT-ZrZNzJVICA.jpeg" /></figure><p><em>Input: official MendTrue synthetic travel-selfie sample used by the backlit-recovery template. Studio/demo image, not a private vacation photo.</em></p><p>The brief was narrow: lift only underexposed areas that still hold evidence, protect the bright sky and water, and refuse to invent detail in fully clipped regions or redraw clothing and expression.</p><figure><img alt="AI backlit-recovery result for the same synthetic sample: lifted recoverable shadows while protecting bright sky and water" src="https://cdn-images-1.medium.com/max/1024/1*zQ2T6G9B95QQWQPZ55Pc_A.jpeg" /></figure><p><em>Result from the current backlit-recovery / shadow-lift path. This is an AI exposure estimate, not a second shutter click at a different time of day.</em></p><h3>I start by asking what is still in the file</h3><p>Before touching “beauty,” I check whether the face is truly empty or merely dark.</p><p>If hair edges, glasses frames, jacket seams, and the path texture are still visible under the haze, the job is selective lift. If a cheek is a pure clipped silhouette with no residual structure, the honest answer is that the model would be guessing — and that guess is not archival evidence.</p><p>For this sample the shadows still held shape. So the test is: can the tool raise the face and path without bleaching the sky into a new postcard.</p><h3>Protect the bright side harder than you “fix” the dark side</h3><p>The bright background is the constraint. Once the sky or water turns milky, or a halo rings the hairline, the edit stopped being recovery and started being a composite look.</p><p>In review I zoomed on:</p><ul><li>hairline and glasses against the bright field — halo or not;</li><li>jacket color and logo-like shapes — still the same garment;</li><li>path stones and horizon line — geometry unchanged;</li><li>water/sky highlights — still clipped where they were, not reinvented with fake clouds;</li><li>expression and age cues — same person, not a remade portrait.</li></ul><p>If any of those move, I treat the result as failed even if the face “looks nicer.”</p><h3>What I refuse to ask in the same pass</h3><p>I do not stack “make it cinematic,” “remove the stranger,” “colorize,” or “sharpen the eyes” on top of a backlight lift. Each of those is a different product decision and a different credit.</p><p>I also do not ask for a golden-hour remake. A remake invents a lighting history the sensor never recorded. The useful question is smaller: recover what the underexposure still contains, and leave clipped regions alone.</p><p>You can repeat the same narrow check on MendTrue’s <a href="https://mendtrue.com/photo-templates/backlit-recovery">backlit scene recovery template</a>: one synthetic sample, one shadow-lift instruction, and a hard rule not to invent a second person or a second sky.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0d08b5f89f64" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Цвет в семейном архиве — это гипотеза, а не найденная плёнка]]></title>
            <link>https://medium.com/@qingwancong/%D1%86%D0%B2%D0%B5%D1%82-%D0%B2-%D1%81%D0%B5%D0%BC%D0%B5%D0%B9%D0%BD%D0%BE%D0%BC-%D0%B0%D1%80%D1%85%D0%B8%D0%B2%D0%B5-%D1%8D%D1%82%D0%BE-%D0%B3%D0%B8%D0%BF%D0%BE%D1%82%D0%B5%D0%B7%D0%B0-%D0%B0-%D0%BD%D0%B5-%D0%BD%D0%B0%D0%B9%D0%B4%D0%B5%D0%BD%D0%BD%D0%B0%D1%8F-%D0%BF%D0%BB%D1%91%D0%BD%D0%BA%D0%B0-5f6d66c5c94c?source=rss-c956225e808e------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f6d66c5c94c</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[photography]]></category>
            <category><![CDATA[family-history]]></category>
            <dc:creator><![CDATA[Qingwancong]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 10:30:37 GMT</pubDate>
            <atom:updated>2026-10-12T10:30:37.700Z</atom:updated>
            <content:encoded><![CDATA[<p>В семейном архиве цвет почти всегда появляется как спор, а не как факт.</p><p>Кто-то уверен, что глаза были серыми. Кто-то помнит платье зелёным. Третий говорит, что «плёнка сама всё покажет», если её «раскрасить». На чёрно-белом отпечатке этих данных нет. Есть яркость, зерно и форма. Цвет, которого не было в файле, модель может только предположить.</p><p>Я не разбирал чужой семейный альбом. Для проверки текущего инструмента MendTrue взят публичный синтетический образец с сайта. Это демонстрационный портрет в винтажном стиле, а не фотография реальной родственницы.</p><figure><img alt="Публичный синтетический образец MendTrue: студийный портрет в винтажном стиле, слева монохром, справа уже показан цветовой вариант" src="https://cdn-images-1.medium.com/max/371/1*7kkwoQF41VtfvxTuwf5upA.jpeg" /></figure><p><em>Входной файл — официальный синтетический sample </em><em>/assets/samples/colorize.jpg. Сайт показывает его как сравнение; я загрузил именно этот файл, не семейный снимок.</em></p><p>Задача была узкой: получить одну сдержанную цветовую интерпретацию, не ремонтируя биографию человека и не превращая портрет в современную глянцевую карточку.</p><figure><img alt="Новый результат AI-колоризации того же синтетического портрета, полученный 2026–09–02 через current colorize preset" src="https://cdn-images-1.medium.com/max/1024/1*eJ2b92PfVP56hz9FdrLnQA.jpeg" /></figure><p><em>Результат от 2026–09–02, preset </em><em>colorize. Это генеративная раскраска, а не найденный оригинальный негатив.</em></p><h3>Сначала сохраняю яркость, цвет добавляю вторым слоем</h3><p>На чёрно-белом портрете уже есть структура, которую опасно ломать: линия волос, серьга, воротник, направление взгляда, зерно бумаги или плёнки.</p><p>Если сразу просить «сделай живую и красивую», модель получает разрешение менять и тон, и геометрию, и кожу. Тогда нельзя понять, что именно произошло: появился цвет или сменилось лицо.</p><p>Поэтому в этой пробе я не смешивал реставрацию царапин, открытие глаз и раскраску. Только хрома. Яркость и кадрирование должны остаться опорой, по которой потом можно спорить с результатом.</p><p>Короткое направление было таким:</p><blockquote><em>Добавь только сдержанную, правдоподобную цветовую интерпретацию. Сохрани лицо, причёску, воротник, зерно и кадр. Не омолаживай и не приукрашивай.</em></blockquote><h3>Цвет губ и глаз — самые скользкие места</h3><p>Именно здесь родственники начинают верить картинке слишком быстро.</p><p>Красный рот на винтажном портрете выглядит «как тогда». Но без записки, второго цветного кадра или свидетельства это только стилистический выбор модели. То же с радужкой: карий, серый или зелёный оттенок может быть убедительным и при этом выдуманным.</p><p>В проверке я смотрел не «стало ли приятнее», а:</p><ul><li>не съехала ли линия челюсти;</li><li>на месте ли серьга и воротник;</li><li>не превратились ли волосы в другой силуэт;</li><li>не исчезло ли зерно так, что кожа стала пластиковой;</li><li>не появился ли современный макияж, которого не было в источнике.</li></ul><p>Если эти опоры держатся, цветовую версию можно оставить как <em>одну из гипотез</em>. Если нет — файл не архив, а другой человек в похожей одежде.</p><h3>Как я бы подписал копию в семейной папке</h3><p>Для важного портрета я бы хранил четыре объекта с понятными именами:</p><ol><li>нетронутый скан или файл;</li><li>при необходимости — отдельно отреставрированный монохром;</li><li>колоризованную копию с пометкой, что цвет сгенерирован;</li><li>короткую запись, какой инструмент и какая формулировка были использованы.</li></ol><p>Так через год никто не спутает гипотезу с единственным оставшимся «оригиналом». Генеративная раскраска может быть аккуратной. Она всё равно не доказывает, какого цвета было платье в день съёмки.</p><p>Проверить такой осторожный прогон можно в <a href="https://mendtrue.com/ru/photo-restoration">странице реставрации старых фото MendTrue</a>: один файл, одно цветовое предположение, обязательное сравнение с источником.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f6d66c5c94c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>