# What are the ethical guidelines for restoring photographs with AI?

colorizethis.io · September 7, 2026

> What Counts as Ethical Photo Restoration in 2026 Ethical photo restoration is the set of practices that govern when, how, and to what extent a damaged...

## What Counts as Ethical Photo Restoration in 2026

Ethical photo restoration is the set of practices that govern when, how, and to what extent a damaged, faded, or monochrome image may be altered while still respecting its status as a historical document. The American Institute for Conservation (AIC) Code of Ethics, originally approved by its Fellows and Professional Associates, frames the central principle: every intervention must be reversible where possible, fully documented, and proportionate to the conservation goal. For AI-assisted work, this translates into three additional duties: disclose that machine learning was involved, preserve the unedited original file as a master, and never fabricate content (faces, text, backgrounds) that did not exist in the source material. The 2024 Nature paper on the Yongle Palace murals and the parallel Nature study on Kizil Grotto murals both used deep-learning optimization and multimodal controlled diffusion models under explicit museological supervision, which demonstrates that the research community accepts AI restoration only when it operates inside established conservation ethics rather than outside them.

**Also worth reading:** [What are the definitive ethical AI photo colorization guidelines for preserving historical accuracy and respecting copyright?](https://colorizethis.io/knowledge/what_are_the_definitive_ethical_ai_photo_colorization_guidelines_for_preserving_historical_accuracy_and_respecting_copyright.php) · [Is it ethical to AI-colorize old family photos? What you should know before restoring black-and-white pictures?](https://colorizethis.io/knowledge/is_it_ethical_to_ai-colorize_old_family_photos_what_you_should_know_before_restoring_black-and-white_pictures.php) · [Can AI colorization provide historically accurate colors for old photographs?](https://colorizethis.io/knowledge/can_ai_colorization_provide_historically_accurate_colors_for_old_photographs.php)

## Why Ethics Became More Pressing After 2018

Before 2018, most photo restoration happened on a light box with cotton swabs and Photoshop curves. Public trust in the process was high because the manipulation was local and visible. That changed when neural colorization tools such as DeOldify (open-sourced mid-2018) and MyHeritage In Color entered consumer hands, followed by Stable Diffusion-based pipelines in 2022. PetaPixel's 2023 feature on AI restoration of Victorian portraits showed how quickly a faded carte-de-visite can be made to look like a modern studio portrait, raising legitimate questions about whether descendants are seeing the ancestor or a statistically average face drawn from millions of training images. The 2018 release of Peter Jackson's They Shall Not Grow Old intensified the debate: Jackson's team rebuilt lip movements and facial detail on 100-year-old 18 fps footage using frame interpolation, and the project drew sustained criticism for crossing the line between restoration and dramatization, as reported by Indiana University News. Photojournalism bodies such as DigitalCustom Photojournalism publish ethics guidelines specifically to keep documentary imagery inside verifiable bounds; the same logic now applies to family and archival photos.

## The Five Pillars of an Acceptable Workflow

A workflow that satisfies most institutional reviewers in 2026 has five pillars. First, capture or scan at the highest lossless resolution available (TIFF, 16-bit, at least 600 ppi for prints). Second, never overwrite the master; all AI runs produce derivative files stored in a versioned folder. Third, annotate every pass with a metadata sidecar that records the model name, version, checkpoint date, prompt (if any), and the operator's initials. Fourth, restrict inpainting and generative fill to regions where the original pixels are physically missing (scratches, mold, emulsion loss) rather than to areas that are merely low contrast. Fifth, retain a side-by-side comparison image and, for any public release, a written justification explaining why each generative change was unavoidable. These five pillars line up with the AIC's broader code of ethics for movable cultural property, which assigns each institution the responsibility to adapt the framework to its own collection.

## Restoration vs Manipulation: Where the Line Sits

The ethical line is usually drawn at whether the change is additive or reconstructive. Reconstructive work replaces pixels that are demonstrably lost using evidence from neighbouring regions of the same image, from a known reference photo of the same subject, or from documented period detail (uniform buttons, wallpaper patterns, hairstyles). Additive work invents details that have no source evidence: a smile where the mouth is blurred, eyes opened where the lids are closed, a background building that may or may not have existed. The Yongle Palace mural project is a textbook example of reconstructive work because the surviving 80% of pigment guided the model on what the missing 20% should resemble. By contrast, the They Shall Not Grow Old team added facial hair, eye colour, and lip motion to soldiers whose faces were at best a 4-pixel smear, which is widely considered additive rather than reconstructive. Generative AI raises the stakes because diffusion models are trained to produce plausible detail, and plausible is not the same as accurate.

## Comparison of Common Restoration Approaches

Different tools carry different ethical weight, and a restorer should pick a method that matches the documentary value of the image.

| Approach | Reversibility | Disclosure Burden | Risk of Fabrication | Best Suited To |
| --- | --- | --- | --- | --- |
| Manual retouch (clone/heal) | Fully reversible | Low (notes only) | Very low | One-of-a-kind heirloom prints |
| Traditional auto-tone (Levels, Curves) | Fully reversible | Low | None | Faded but intact originals |
| Diffusion-based colorization (DeOldify, MyHeritage) | Reversible via layers | Medium (model + version) | Low for monochrome, medium for faces | Black-and-white portraits |
| Generative fill / inpainting (Stable Diffusion, Adobe Firefly) | Reversible only if masks saved | High (prompt + seed) | High if scope is large | Genuine scratch and tear repair |
| Hybrid pipelines (manual mask + diffusion inpaint) | Reversible | High | Medium | Damaged historical photographs |
| Pure creative reinterpretation | Not intended to be reversible | Must be labelled art | Effectively certain | Decorative prints, not archives |

The table makes the trade-off explicit: the higher the fabrication risk, the heavier the disclosure and documentation burden, and the more carefully the result must be labelled so viewers cannot mistake it for a faithful reproduction.

## Practical Steps for a Family Archivist or Small Museum

Most people restoring a family archive in 2026 are not professional conservators, but the same scaffolding applies at a smaller scale. Start by making two scans of the original: one at archival quality and a working copy at 300 ppi. Open the working copy in a non-destructive editor (Adobe Photoshop with layers locked, GIMP 2.10, or Affinity Photo 2) and confine every adjustment to a new layer. For monochrome images, run a colorization model only on a duplicate layer, then toggle that layer off to confirm the original has not been touched. For scratch and tear repair, mask the damaged region, run an inpainting model with a fixed seed, and export the result as a separate file named original_damaged.tif and restored_v1.tif. Save the AI prompt, model version, and seed in a plain-text manifest alongside the images. If the photograph will be published, attached to a genealogy site, or printed in a family history book, add an unobtrusive caption such as "AI-assisted restoration, generative fill applied to damaged regions only, original held by the family." This level of disclosure is consistent with the spirit of the AIC Code of Ethics and with the photojournalism norms published by DigitalCustom Photojournalism.

## Common Mistakes That Compromise Ethics

Five errors appear repeatedly in poorly governed restoration projects. The first is using a single AI pass as a final master without keeping the original scan, which destroys the evidentiary baseline. The second is letting a diffusion model fill in faces from low-resolution regions, producing an output that is technically sharp but statistically invented. The third is failing to record the model and version, so a 2026 restoration cannot be reproduced or audited three years later when the underlying model has been retrained. The fourth is applying the same colour palette to every image in a batch, producing an unnatural visual uniformity that misleads viewers about period reality. The fifth is sharing restored images on social media without any disclosure, allowing the work to circulate as an unedited original. All five are avoidable with a short written protocol.

## When to Seek a Professional Conservator Instead

DIY AI restoration is appropriate for personal family photographs, school projects, and casual social media posts. It is not appropriate when the image has monetary, legal, or evidential value. Photographs intended for sale at auction, evidence in a legal proceeding, submission to a national archive, or display in a museum exhibition should be handled by a professional conservator, ideally one who is a Professional Associate of the AIC. The same threshold applies to daguerreotypes, ambrotypes, and other one-of-a-kind processes whose surfaces cannot be re-scanned if the original is damaged. Professional fees for a single 8x10 archival restoration in 2026 typically range from $180 for a basic digital clean to $1,200 for a multi-day intervention involving microscopy and chemical stabilization, with AI-assisted steps billed at $60 to $150 per hour on top. Budget roughly two to four weeks of turnaround for routine jobs and two to six months for museum-grade work that requires peer review of the proposed treatment.

## The Legal and Privacy Layer in 2026

Ethics and law are not the same thing, but in 2026 they increasingly overlap. The EU AI Act, in force since August 2024 and in its enforcement phase by mid-2026, classifies AI systems that generate or manipulate biometric data as high-risk when used in employment, education, or law enforcement contexts, but the underlying GDPR still governs any processing of identifiable faces. In the United States, no federal law yet directly regulates AI photo restoration, but the right of publicity statutes in states such as California, New York, and Tennessee protect living individuals from commercial reuse of their likeness, which can include colourised reproductions of identifiable minors. Canadian and Australian privacy frameworks treat a recognisable face as personal data, so publishing a colourised version of a stranger's old photograph without consent can trigger complaints to the privacy commissioner. Restorationists should keep copies of consent forms, document the lawful basis for processing, and avoid circulating AI-restored images of living private individuals on public platforms without a clear purpose.

## How This Connects to AI Colorization Tools Like colorizethis.io

AI colorization sits inside the same ethical frame as AI restoration, and a tool such as colorizethis.io inherits the duties above. A responsible colorizer should keep the user's original upload unchanged, return a separate colourised derivative rather than overwriting, name the model and version in any exported metadata, and refrain from inventing fine facial detail when the source is below roughly 256x256 effective resolution. It should also display an unobtrusive badge on the exported image (or embed it in the file's IPTC or XMP block) stating that AI colourisation was applied, and offer a side-by-side preview so the user can see exactly what changed. None of these features are expensive to add, and each one closes a gap that the AIC Code, the EU AI Act, and basic photojournalism norms have flagged repeatedly since 2022. The cumulative effect is that AI colorization stops being a magic trick and becomes a documented conservation act, which is the only position a public-facing service can defend in 2026.

## Quick answers

### Is AI photo restoration considered ethical?

AI photo restoration is ethical when the original is preserved, every intervention is documented with model name and date, generative fills are limited to demonstrably missing regions, and the result is labelled as AI-assisted. It becomes unethical when faces or backgrounds are invented in areas where the source is merely low contrast, or when the original scan is overwritten.

### Do I need to disclose that a photo was restored with AI?

Yes for any public or commercial use. The AIC Code of Ethics, the EU AI Act's transparency rules, and most photojournalism guidelines require that AI-assisted images carry a visible caption or embedded metadata stating the technique, the model, and the date. Personal family use is the main exception where disclosure is optional.

### What resolution is too low for safe AI face restoration?

Faces below roughly 256x256 effective pixels, or roughly 50x70 pixels per face, are widely considered too low for credible facial reconstruction because diffusion models begin to average features rather than recover them. Below that threshold, conservators usually recommend leaving the face unrestored and documenting the limitation in the caption.

### How much does professional photo restoration cost in 2026?

Standard archival digital cleaning of an 8x10 print runs $180 to $400 in 2026. AI-assisted work adds $60 to $150 per hour. Full museum-grade restoration with chemical stabilization and peer-reviewed treatment plans typically ranges from $900 to $3,500 per object, with turnaround of two weeks to six months depending on queue.

### Can AI restoration be reversed if a better model comes out?

Only if the original scan and every layered working file have been preserved. Because diffusion models are non-deterministic, re-running the same prompt and seed is reproducible, but a higher-resolution or differently trained future model requires the untouched original to be re-processed from scratch. Reversibility is the central reason institutional protocols forbid overwriting the master file.

Canonical: https://colorizethis.io/knowledge/what_are_the_ethical_guidelines_for_restoring_photographs_with_ai.php
Markdown: https://colorizethis.io/knowledge/what_are_the_ethical_guidelines_for_restoring_photographs_with_ai.php/index.md