Industries / Retail, hospitality and personal services / Photography
Editing
This is the work done to a picture after it is taken and before it is delivered: combining and retouching, setting colour and tone, removing what should not be there, and labelling pictures whose people have been changed.
7
Milestones achieved, from 1857 to 2023.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1857
One print from many negatives
Oscar Gustav Rejlander made The Two Ways of Life from more than 30 separate negatives, each exposed on its own and then printed together as one picture, so large that it needed two sheets of paper joined together. By the Victoria and Albert Museum's account, this "combination printing" showed that "a photograph could be composed in the manner of painting".
Credit
Oscar Gustav Rejlander
1950s
One face sets print colour
From the mid-1950s Kodak sent independent photo labs a kit of prints and unexposed negatives of a studio model, Shirley Page, so that they could set the colour and density of its new S5 printers; the reference card kept the name "Shirley" whoever posed for it. For decades the models were white and often labelled "normal". Lorna Roth of Concordia University, who researched the cards, says the film's range favoured white skin, so pictures of Black and white people together came out partly under- or over-exposed. In the 1970s the Kodak photographer Jim Lyon began testing film on Black models, and Kodak later made reference cards with Black, Asian, Latina and white models, one of them dated 1996.
LessonA colour reference built around one light skin tone carried its bias into every print matched to it.
Credit
Kodak, whose photo tech division was run by Ray DeMoulin; Jim Lyon
1982
Pyramids moved for a cover
For its February 1982 cover, National Geographic had Gordon Gahan's horizontal photograph of a camel train at Giza altered by computer to fit the vertical cover, which moved the pyramids closer together. In 1985 the magazine's director of engraving and printing, William W. Smith, defended it, saying the photographer could have made the same picture by moving a few feet. In 2016 the editor, Susan Goldberg, called the outcry "deserved", wrote that altering a photo is never acceptable at the magazine, and said it now asks photographers on assignment for their raw files, which hold the data straight from the camera's sensor.
LessonAn edit the editors thought harmless cost readers' trust, and the remedy was a rule backed by a check against the camera's own file.
Credit
National Geographic; photograph by Gordon Gahan
1990
Photoshop ships for the Mac
Thomas Knoll, a PhD student at the University of Michigan, wrote a program in 1987 to show and change digital images, and his brother John, at the film effects company Industrial Light & Magic, used it to edit photos. The scanner maker Barneyscan bundled about 200 copies of an early version, Adobe licensed it in April 1989, and Photoshop 1.0 for the Apple Macintosh shipped early in 1990. More than 3 million copies sold in the next ten years, by the Computer History Museum's account.
Credit
Thomas and John Knoll; Adobe, urged on by its art director Russell Brown
2007
Whole shoots in one program
Adobe said in January 2007 that Photoshop Lightroom 1.0, a program for sorting and adjusting large numbers of digital photos, would ship in mid-February, competing with Apple's Aperture. Its Develop module kept several versions of a photo as "virtual copies" without saving extra files on disk, by TidBITS's account. When MIT and Adobe researchers later built the FiveK set, 5,000 raw photos each adjusted by five photography students, the students did the work in Lightroom.
Credit
Adobe
2017
Labels for slimmed models
From 1 October 2017, French law requires the words "photographie retouchée" (retouched photograph) on commercial photos of models whose body outline has been slimmed or thickened with image software, in the press, on posters and online, though not on television; the advertiser must check with its photographers, and the fine is €37,500, which can rise to 30% of the advertising spend. Norway went further from 1 July 2022: a standard label is required on any advertising in which a person's body shape, size or skin, face included, has been changed by retouching, filters or lenses, so that removing a single pimple triggers it, and the rule covers advertising posts on Instagram and TikTok and stock images.
Credit
the French Parliament, through the health law of 26 January 2016 and decree 2017-738; Norway's Marketing Control Act, enforced by the Norwegian Consumer Authority
2023
Fill by typing a request
On 23 May 2023 Adobe released Generative Fill in a test version of Photoshop: the user selects part of a picture and types a short request, and Adobe's Firefly image model adds, extends or removes content there. It followed Content-Aware Fill, which filled a selection with patches copied from elsewhere in the same picture using PatchMatch, a fast patch-matching method published in 2009 by Connelly Barnes, Eli Shechtman, Adam Finkelstein and Dan B Goldman; Barnes's 2011 Princeton thesis says Photoshop CS5 included it. By October 2025 Generative Fill was generally available and offered other companies' models as well, such as Google's Gemini 2.5 Flash Image and Black Forest Labs' FLUX.1 Kontext.
news.adobe.comgfx.cs.princeton.educs.princeton.edunews.adobe.com
Credit
Adobe; for PatchMatch, Barnes, Shechtman, Finkelstein and Goldman
Watch for
Edits made by asking
Google says Google Photos lets people describe an edit in words or by voice, such as "remove the cars in the background" or "restore this old photo", and its Gemini models carry it out; the feature came first to Pixel 10 phones in the US. Adobe opened a public test of an AI Assistant in Photoshop on the web and on phones that applies edits described in words, such as removing distractions, changing backgrounds or adjusting colour, or lists the steps for the user to follow.
Achieved whenAn outside group publishes, for a fixed set of photos and written requests, how often each tool does what was asked without unrequested changes to faces, skin or other content.
Claimed by
Google, August 2025; Adobe, March 2026
Retouching networks tint skin
Choulhyouc Lee reports that a small retouching network, with 6,825 parameters and trained by damaging good photos and learning to repair them, made skin greener than the original in 52 of 70 portraits, while PSNR and SSIM, the usual whole-picture similarity scores, barely moved. A training penalty that looks only at skin cut the tint to 1 of 70 portraits, and the same networks trained on the MIT-Adobe FiveK experts' edits were close to neutral.
Achieved whenAnother group reproduces the tint and the fix on other retouching models, or a maker of a retouching product publishes a test for it.
Claimed by
Choulhyouc Lee, in a paper accepted to the ACCV 2026 computer vision conference, November 2026
Open
01
Skin tone after AI retouching
A public test of the AI retouching tools that portrait and headshot photographers use, measuring how much each one changes skin colour. Photograph consenting adult volunteers across the 10 shades of the Monk Skin Tone Scale, under one fixed light with a colour chart in the frame. Run each tool at its default and strongest settings, then measure, on skin areas only, the change in lightness and hue and the overall colour difference between the photo before and after.
The briefCredit · unclaimed02
Spot slimmed bodies without originals
A detector that flags body-slimming edits in a single photo, and a public test set that scores it by the size of the edit. Consenting adult models are photographed clothed; retouchers narrow waists, arms or thighs with Photoshop's Liquify tool by measured amounts, such as 2%, 5% and 10% of the width, and unedited frames are kept alongside. The detector is scored on the smallest edit it catches and on how often it flags unedited photos.
The briefCredit · unclaimed03
Restoration checked against originals
A test of AI old-photo restoration in which the undamaged original is known. Print public-domain portraits, or portraits of consenting adult volunteers, age the prints by fading in sunlight, folding, scratching and water staining, scan them, and restore each scan with several tools. Compare each result with the original digital file: whether the face is still the same person by a face-recognition score, how far the colours drift, and what detail blind raters mark as invented.
The briefCredit · unclaimed
Open 01 · the brief
Skin tone after AI retouching
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of the AI retouching tools that portrait and headshot photographers use, measuring how much each one changes skin colour. Photograph consenting adult volunteers across the 10 shades of the Monk Skin Tone Scale, under one fixed light with a colour chart in the frame. Run each tool at its default and strongest settings, then measure, on skin areas only, the change in lightness and hue and the overall colour difference between the photo before and after.
Why it is still open
Piera Riccio and Nuria Oliver (2022) studied eight Instagram augmented-reality filters, which change a face live in the camera app, applied to the FairFace photo set, not the retouching tools sold to photographers. Lee (2026) found a green tint in research networks, not in commercial products. The one hands-on test of a commercial tool found, PetaPixel's November 2026 article on Reblum, is sponsored by Reblum, written by a Reblum partner and gives no measurements.
Built with what exists
A camera that shoots raw files; a colour chart such as the Calibrite ColorChecker; trial versions of retouching tools such as Adobe Lightroom, Evoto, Retouch4me and Reblum; Google's open Monk Skin Tone Scale to sort volunteers by shade; Google's MediaPipe or another open face-parsing model to mask skin; the open-source colour-science Python library for CIELAB values (standard coordinates for lightness, red-green and yellow-blue) and CIEDE2000 colour differences.
First step
Photograph ten volunteers who have signed a consent form, under one light with a colour chart in the frame, run two tools, and measure the change on their skin.
Done looks like
At least 100 consenting adults, at least 8 per Monk shade, at least four tools at two settings each, and a public table of the median change in lightness and CIEDE2000 per tool and shade, with the three darkest shades set beside the three lightest, and code that reruns the measurements on new photos.
Open 02 · the brief
Spot slimmed bodies without originals
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A detector that flags body-slimming edits in a single photo, and a public test set that scores it by the size of the edit. Consenting adult models are photographed clothed; retouchers narrow waists, arms or thighs with Photoshop's Liquify tool by measured amounts, such as 2%, 5% and 10% of the width, and unedited frames are kept alongside. The detector is scored on the smallest edit it catches and on how often it flags unedited photos.
Why it is still open
France requires a label on commercial photos whose body outline was slimmed or thickened, and Norway on any change to body shape, size or skin, but both rest on the advertiser asking the photographer. Eric Kee and Hany Farid's 2011 retouching score needs both the original and the retouched photo. Sheng-Yu Wang and colleagues (2019) detect Photoshop warping of faces only, and RetouchingFFHQ (2023) covers face retouching. BR-5K (2022) holds 5,000 Unsplash portraits with bodies reshaped by professionals, built for making such edits rather than catching them, and is released for academic, non-commercial use after a signed agreement. An arXiv search in November 2026 found no detector for reshaped bodies.
Built with what exists
BR-5K under its access agreement; Photoshop's Liquify tool driven by scripts, the method Wang and colleagues used to make warped faces; an open pose model such as OpenPose to measure body widths before and after; PyTorch for the detector.
First step
Apply for BR-5K access and, on 100 of its pairs, measure how much the professional edits narrow waists and arms, to choose the edit sizes for the test.
Done looks like
A test set of at least 2,000 photos of at least 50 consenting adults, half unedited, with the detection rate published for each edit size and the share of unedited photos wrongly flagged, and the detector and test set released for others to rerun.
Open 03 · the brief
Restoration checked against originals
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test of AI old-photo restoration in which the undamaged original is known. Print public-domain portraits, or portraits of consenting adult volunteers, age the prints by fading in sunlight, folding, scratching and water staining, scan them, and restore each scan with several tools. Compare each result with the original digital file: whether the face is still the same person by a face-recognition score, how far the colours drift, and what detail blind raters mark as invented.
Why it is still open
Google says Photos will "restore this old photo" when asked. The real old-photo set released with Pik-Fix (2022) pairs each damaged photo with a version "manually restored by PhotoShop experts", a human restoration rather than the original, and Li and colleagues (2025), who built VintageFace, a benchmark of 300 real old face photos, write that this kind of damage makes large paired sets hard to build. Justin Norman and Hany Farid (2024) measured whether AI enhancement keeps a face's identity, but on photos blurred and shrunk by computer for forensic face recognition, not on aged prints. PetaPixel's March 2026 opinion piece showed ON1's Restore AI changing a man's face, without measurements.
blog.googlearxiv.orgarxiv.orgfarid.berkeley.edupetapixel.com
Built with what exists
Public-domain portraits, such as Library of Congress photos marked "no known restrictions"; an inkjet printer or a photo lab; a sunny window, a craft knife and water for ageing; a flatbed scanner; restoration tools such as Google Photos, Photoshop's Photo Restoration neural filter, MyHeritage, Remini and ON1 Restore AI; ArcFace face matching through the open InsightFace library; the colour-science Python library for colour differences.
First step
Print 20 public-domain portraits twice, keep one copy of each in the dark and tape the other to a sunny window for four weeks, scanning both every week.
Done looks like
At least 200 aged prints across at least four kinds of damage, each restored by at least five tools, with a public table of face-match scores against the original, colour error, and the share of restorations that raters judged to show a different person, and the scans released under an open licence.