Old photographs fade, crack, stain and arrive as tiny scans. Generative and discriminative image models can remove defects, upscale and add plausible colour in minutes instead of hours of manual retouching. The output is a reconstruction, not a record: faces, textures and colours are partly invented, so archives and publishers must label results and keep the original untouched.
Problem & value
- Today: restoration is done by retouchers in Photoshop (clone stamp, healing, manual colour layers), costing from tens of minutes to many hours per image; families and small archives often leave shoeboxes of photos unscanned or unrepaired.
- AI value: batch cleaning of dust, scratches and noise, super-resolution of small scans, and one-click colourisation give a good first result; a human retoucher only finishes the hard images. This makes digitisation projects (local history, genealogy, museum collections, newsroom archives) affordable and increases reuse of archive material.
- No reliable public benchmark gives "hours saved" figures across providers; measure it on a sample of your own material.
Variants
- Private / genealogy: family photos via apps or consumer web tools; privacy of relatives matters.
- Archives, libraries, museums (cultural heritage): batch restoration with strict provenance; often only non-destructive enhancement is acceptable and colourisation is labelled as interpretation.
- Media & documentaries: colourised stills and film frames for editorial use (video: see GEN-012).
- Genealogy and memorial services: restored portraits for obituaries, gravestones, books.
- Heritage science / forensics-adjacent: not suitable as evidence; generative output must never be used as proof of appearance.
How it works
- scan (TIFF, >=600 dpi for prints, keep the master unchanged)
- analysis (damage mask: scratches/tears/stains; face and text regions detected)
- inpainting of damaged areas (diffusion or GAN inpainting, masks drawn or predicted)
- denoise + deblur + super-resolution (Real-ESRGAN / SwinIR class)
- face restoration (optional, identity risk)
- colourisation (DDColor / diffusion-based, optional colour hints from metadata)
- human review against the original (faces, uniforms, flags, signage)
- export with edit log + provenance label (C2PA / IPTC "digitally altered")
- Restoration models: GAN or diffusion models trained on synthetic degradations. Instruction-based image editors (a prompt such as "remove scratches and restore, keep identity") work as a general alternative to specialised tools.
- Face restoration: models such as GFPGAN and CodeFormer use face priors; they sharpen faces but can substitute a different plausible face, which is a serious problem for identifiable ancestors or historical persons.
- Colourisation: predicts the most likely colour from context; for ambiguous objects (clothing, cars, flags) the result is a guess. Adding a vision-language model to describe the scene helps check historical plausibility.
- Human-in-the-loop: compare with the original, correct colours using documentary evidence, approve before publication.
Data preparation
The input is image material: prints, negatives, slides and glass plates in boxes, albums or archive stores, plus already scanned TIFF/JPEG files on file shares or in a DMS or media database. In an SMB or small archive this often means physical material that has never been digitised, and scanned files with inconsistent names, no dates and no rights information.
Typical gaps: low scan resolution, JPEG-compressed copies instead of masters, duplicates, missing metadata (date, place, persons), unknown copyright and unclear consent of depicted people, and no separation between original and edited version.
Preparation steps: (1) scan at 600 dpi or more as uncompressed TIFF, keep this master untouched and read-only; (2) deduplicate and give files stable IDs; (3) record minimal metadata and rights status in a spreadsheet or the DAM; (4) add owner hints (clothing colours, uniforms, place) that guide colourisation; (5) agree on a naming scheme for restored derivatives and an edit log. Tools: flatbed or book scanner, Paperless-ngx or a DAM for storage, ExifTool for metadata.
Rough share of effort: about 30 to 40 percent, mostly scanning and metadata; the AI step itself takes minutes per image. See PREP-006 for scanning, PREP-010 for metadata and tagging and PREP-009 for restructuring file shares.
Implementation paths
Which path for whom
- smb: Path A, because mature restoration and colourisation services exist at per-image or small subscription prices and the process is standard.
- enterprise: Path B, because archives and media houses must plug restoration into their DAM and provenance workflow and process volumes where per-seat tools get expensive.
- developer: Path C, because open models (restoration, upscaling, colourisation) run on HW2 and can be wrapped into own tools without per-image fees.
- private: Path C, because local apps such as Upscayl or ComfyUI keep family photos on the own computer and cost nothing per image.
- public: Path C, because museums, archives and municipalities often have procurement rules and sensitive holdings (people, victims, minors) that favour sovereign or local operation.
Path A — Buy (recommended for SMBs)
- Consumer/prosumer tools: MyHeritage (Photo Enhancer, Colorize; Israel-based, EU users apply GDPR), Remini, VanceAI, Topaz Photo AI (desktop, local processing), Adobe Photoshop Neural Filters (Photo Restoration, Colorize) and Firefly, Google Photos / Pixel tools for private use, Hotpot.ai.
- Archive-grade/service providers: specialised digitisation bureaus offering AI-assisted restoration (check EU processing).
- Choose A when: volume is low to medium and the images are not highly sensitive. Prefer local desktop tools for sensitive material.
Path B — Build on cloud APIs
- Image-edit APIs: instruction-based editors such as Black Forest Labs FLUX.1 Kontext (Germany; hosted API, EU-hosting via partners), OpenAI image editing API, Google Gemini image editing ("Nano Banana"), Stability AI image services, Adobe Firefly Services. Stronger on global edits, weaker on exact fidelity: always diff against the original.
- Super-resolution endpoints: Replicate, fal.ai or cloud marketplaces offering Real-ESRGAN/Clarity-type upscalers.
- Orchestration: n8n or a small queue service: upload → API → review UI → archive. Pick EU regions and sign a DPA when portraits of identifiable persons are processed.
Path C — Self-host (open source)
- Models/tools: Real-ESRGAN, SwinIR, GFPGAN, CodeFormer (check licence: some weights are non-commercial), DDColor (colourisation), DeOldify (older), LaMa for inpainting, FLUX.1 Kontext [dev] (open weights, non-commercial licence unless licensed), Qwen-Image-Edit (Apache 2.0), Stable Diffusion/SDXL inpainting pipelines in ComfyUI; BOPBTL ("Bringing Old Photos Back to Life") for combined restoration research.
- Stack: ComfyUI or diffusers, a simple web review UI, checksum-based storage of originals.
- Hardware: HW2 (one 16–32 GB GPU) is enough for batch upscaling/colourisation; larger diffusion editors run comfortably on 24 GB with quantisation. Data never leaves the premises: best for municipal archives and museums with GDPR concerns.
Path D — Custom training
Worthwhile only for special material (glass plates, specific film stocks, a damage type typical of one collection). Needs paired clean/degraded examples, usually synthesised by degrading clean scans of the same collection; fine-tune a restoration model or add a LoRA to an image editor. Colour accuracy can improve if reference colour photos from the same era and region exist.
State of the art (2026)
- Instruction-based editors (FLUX.1 Kontext, Qwen-Image-Edit, GPT-image, Gemini image models) now do restoration and colourisation in one prompt with good identity retention, replacing multi-tool chains for many users.
- Specialised models (Real-ESRGAN, DDColor, CodeFormer) remain faster, cheaper and more controllable for bulk work.
- Limits and failure modes: hallucinated detail in faces, text, uniforms and medals; "over-smoothed" plastic skin; wrong or anachronistic colours; inconsistent colours across a series; loss of film grain that is part of the document; biased skin-tone colourisation. PSNR/SSIM scores do not reflect fidelity of historical content, so evaluation needs expert visual review.
- Provenance: C2PA content credentials and IPTC digital-source-type values allow marking "edited with AI"; support is growing in Adobe tools and cameras.
Shared resources & combinations
- Same image-gen stack and GPU as GEN-011 (retouching, upscaling, inpainting), GEN-020 (synthetic data) and GEN-026 (portraits); video frames via GEN-012.
- Archive pipelines: KNOW-019 (heritage collection discovery) and KNOW-016 (multimodal search): restoration, captioning with a VLM and indexing share the digitisation workflow.
- Segmentation and VLM captioning help generate damage masks and descriptive metadata.
Effort, cost & KPIs
- Path A: hours; subscription or per-image credit pricing.
- Path B: 2–6 weeks for an intake→review→archive workflow; cost per image is mainly API fees.
- Path C: 1–2 months; one GPU workstation plus a maintainer.
- Path D: quarters, only for collection-specific needs.
- KPIs: share of images accepted without manual rework, retoucher minutes per image, reviewer-rated fidelity (blind A/B against original), share of images with provenance label, digitised-and-published images per month.
People & change
Involve management (budget, rights policy) and the department lead (archive, communication or marketing) at the start, then IT for storage and accounts, and the end users who currently retouch or catalogue photos. Bring in the data protection officer early when photos show identifiable people, especially minors or historic persons with living relatives; an AVV (Art. 28 GDPR) is needed for every cloud service. A DPIA is normally not required for restoring photos but check it for large face-recognition or tagging add-ons.
Works council: normally no co-determination, because the tool does not monitor staff behaviour or performance (§87(1) no. 6 BetrVG); still inform the council when retouching work processes change (§90(1) no. 3). In public bodies involve the staff council if applicable.
AI literacy: offer a short training on what restoration invents versus restores, and on labelling (Art. 4 AI Act requires measures that support literacy). Communicate openly that retouchers keep control: they review every result, can reject colourisation and decide what is published. Start with a pilot group of two or three people and 50 to 100 images. Typical concern: "AI replaces the retoucher". Answer: it removes routine repair so that experts handle difficult cases and quality control.
See GUIDE-008 for change management, GUIDE-005 for data protection and GUIDE-006 for AI literacy.
Risks, regulation & compliance
- EU AI Act: generally limited risk via Art. 50 transparency: deployers of AI-generated or manipulated image content that is a deep fake must disclose it; artistic, creative or satirical works have a lighter disclosure duty. Providers of generative systems must machine-mark outputs. Plain restoration of non-persons is minimal risk. Obligations under Art. 50 apply from August 2026 (check the current state of the Digital Omnibus amendments). Restoration used inside a prohibited or high-risk system (e.g. identification) is out of scope here.
- GDPR: photos of identifiable living people are personal data; deceased persons are outside GDPR but national law and personality rights (post-mortem protection) may apply. Genealogy sites commonly process third-party faces: use a DPA, EU hosting, and avoid providers that train on uploads. Face-restoration is not biometric identification, but do not combine it with face matching.
- Copyright: restoration does not create or transfer rights; check rights of the photographer (in the EU, term typically life + 70 years; simple photographs may have shorter protection under national law) and the licence of any model weights.
- Integrity/disinformation: colourised or "enhanced" images can be mistaken for authentic documents. Mitigate with visible labels, preserved originals, edit logs and a rule that restored images are never used as evidence.
- Mitigations: keep master scans immutable, label outputs, review faces and symbols, avoid training-on-upload providers.
Getting started
- Digitise a pilot set (50 photos) at suitable resolution; store untouched masters with checksums.
- Run each image through one Path A tool and one open-source chain (Real-ESRGAN + DDColor); compare side by side.
- Define a policy: what may be changed (dust, colour) and what not (faces, text, symbols), and the label wording.
- Add provenance metadata (IPTC digital source type / C2PA) in the export step.
- Measure retoucher time saved and acceptance rate; only then scale up or self-host.
Sources
- Regulation (EU) 2024/1689 (AI Act), Art. 50 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Real-ESRGAN — https://github.com/xinntao/Real-ESRGAN
- DDColor (ICCV 2023) — https://github.com/piddnad/DDColor
- CodeFormer — https://github.com/sczhou/CodeFormer
- Black Forest Labs, FLUX.1 Kontext — https://bfl.ai/models/flux-kontext
- Wan et al., "Bringing Old Photos Back to Life" — https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life
- C2PA specification — https://c2pa.org/specifications/specifications/2.1/index.html
- Adobe, Photoshop Neural Filters (Photo Restoration, Colorize) — https://helpx.adobe.com/photoshop/using/neural-filters.html
- GDPR, Regulation (EU) 2016/679 — https://eur-lex.europa.eu/eli/reg/2016/679/oj