AI Image Restoration: How to Restore Old Photos with AI (Tools, Workflows, and Pitfalls)

A century of family history can live on a brittle 4x6 print. AI photo restoration now means you don’t need a darkroom or months of training to bring that image back to life—you need a clear workflow, the right tools, and a light touch. This guide moves you from research to execution with practical steps, comparisons, and decision frameworks you can use today.

Quick Answer

• If you want the fastest one-click result: try a reputable desktop tool like Topaz Photo AI or a privacy-aware mobile app like Google Photos’ “Photo Scan” for capture and Remini (with an offline export for sensitive images). Expect good but sometimes overly smooth results.
• If you need museum-grade control and privacy: scan at high resolution, work locally with non-destructive edits, and use a modular stack: Photoshop (healing/inpainting + Generative Fill), Lightroom (tone), plus local AI like Real-ESRGAN for upscaling and GFPGAN/CodeFormer for gentle face restoration.
• If you prefer open-source end‑to‑end: scan well, then use GIMP or Krita for manual cleanup, Real-ESRGAN for upscaling, Stable Diffusion (inpainting) with a local UI (e.g., ComfyUI/InvokeAI) for tear repair, and optional CodeFormer for faces.
• General order of operations: digitize right → deskew/crop → dust/scratch cleanup → tear/inpainting → denoise/deblur → tone and color correction → optional colorization → gentle face restore → upscaling → print/export.

What AI Photo Restoration Can—and Can’t—Do

AI excels at automating tedious cleanup and reconstructing plausible detail. But “plausible” isn’t always “true.” Decide early whether your goal is:

• Preservation: faithful to the original, blemishes removed, tones corrected, no invented content.
• Presentation: a pleasant image for sharing, even if AI fills gaps.
• Reconstruction: aggressive restoration where missing areas are rebuilt; clearly label edits.

Keep a non-destructive pipeline and save lossless masters. When AI must invent content (e.g., missing background behind a tear), keep a version history and document changes.

Capture Comes First: Scan or Photograph the Print Properly

Garbage in, garbage out applies doubly to restoration. Most “AI problems” begin with poor digitization. Before software, nail your capture.

Best practices for flatbed scanning

• Resolution: For small prints (2x3 to 4x6), scan at 600–1200 dpi; for larger prints (5x7+), 600 dpi is usually enough. If you plan large reprints, err higher.
• Bit depth: Scan at 16‑bit per channel if your scanner allows (gives more latitude for tonal recovery). Save to TIFF.
• Color management: Use your scanner’s color profile if available. Avoid auto color “fixes” in the scan driver.
• Cleanliness: Gently dust the print with a soft, lint‑free cloth or blower. Never use liquids on the print surface.
• Alignment: Place prints square and flat; use scanner guides. Multiple scans for curled prints may be needed.

Photographing prints with a camera or phone

• Lighting: Use diffuse, even light. A bright window with indirect light or two soft lamps at 45° angles works. Avoid glare and hotspots.
• Polarization (optional): If you have a polarizing filter, it helps tame gloss and reflections.
• Stabilize: Use a tripod or prop your phone. Enable gridlines and align edges to avoid keystone distortion.
• Capture format: Shoot RAW on cameras; on phones, use the highest resolution and consider apps that capture RAW (e.g., Halide on iOS).
• Apps: Google PhotoScan reduces glare via multiple frames. For irreplaceable prints, a simple copy stand setup beats handheld.

Handling negatives and slides

• Dedicated film scanners yield better detail than flatbeds for 35mm. If unavailable, use a macro lens and a uniformly lit light panel. Capture RAW and expose to preserve highlights.

Choosing Your Restoration Approach

The right tool depends on your constraints: time, budget, privacy, and desired control. Here’s a practical comparison framework.

Decision shortcut:

• Need local/offline due to privacy? Choose pro editor stack or open-source local.
• Batch a shoebox of photos quickly? Desktop all-in-one with gentle settings.
• Rebuild a torn heritage portrait? Pro editor + local diffusion inpainting.
• Post to social today with minimal effort? Mobile app—with caution about over-enhancement.

Core Restoration Workflow (Step‑by‑Step)

A consistent order prevents fixing one problem only to reintroduce another. The following workflow works for most old photos, whether black‑and‑white or faded color.

1) Prepare the digital master

• Import your scan/shot. Save a Master_Original.tif in a dated folder.
• Deskew and crop conservatively; keep a small border so you don’t cut off edge detail.
• Convert to a wide working color space (e.g., Adobe RGB) if editing in 16‑bit; for simpler pipelines, sRGB is fine. Keep a non-destructive setup: Photoshop Smart Objects or Lightroom.

2) Dust, specks, and minor scratches

• Quick pass: Use Photoshop’s Dust & Scratches filter with a low radius and threshold; mask in only where needed.
• Manual cleanup: Spot Healing Brush or Clone Stamp at 100–200% zoom. In GIMP/Krita, use Heal/Clone.
• AI assist: For batches, some tools have defect detection. Keep it subtle—avoid plastic textures.

3) Tears, missing edges, and large cracks (Inpainting)

• Localize the problem with a rough mask. Don’t include intact facial features in the mask if you can avoid it.
• Photoshop Generative Fill: Prompt with brief factual guidance: “Reconstruct paper texture matching background wall” or “Continue jacket fabric pattern.” Keep prompts literal, not creative.
• Stable Diffusion inpainting (local UI): Use a tile‑friendly model and modest denoise strength. ControlNet (lineart/canny) can preserve structure. Keep the mask tight and feather lightly.
• Iterate at low strength. Merge results on a layer you can opacity-blend.

4) Denoise and deblur—gently

• For grainy scans or high‑ISO camera captures: Try a specialized denoiser (Topaz Photo AI, Lightroom Denoise) at low to medium strength.
• For motion blur, AI can help but be cautious. Over‑sharpening introduces halos and false detail. Consider a small-radius Unsharp Mask or a dedicated AI sharpen tool sparingly.

5) Tone and contrast recovery

• Use Curves for global tonal corrections. Aim for black and white points that respect the paper’s dynamic range.
• Local dodge and burn on a 50% gray overlay layer for gentle balancing.
• For faded black‑and‑white prints, convert to monochrome, then adjust with Curves and a subtle clarity bump.

6) Color correction and restoration

• Old color prints often have a uniform color cast (yellow/magenta). Start with White Balance: click a neutral area if available.
• Use HSL to tame shifts (e.g., oversaturated reds in 1970s prints). Beware of clipping channels.
• Advanced: Use selective color or channel mixer to rebalance. Work in 16‑bit to avoid banding.

7) Optional colorization

• If your original is black‑and‑white and you choose to colorize, treat it as an interpretive step.
• Tools: Photoshop Neural Colorize (beta), DeOldify forks, or Stable Diffusion with reference images. Keep skin tones realistic and clothing era‑appropriate.
• Document that the image is colorized and retain the monochrome master.

8) Face restoration—only if needed

• GFPGAN/CodeFormer can gently reconstruct facial features from small or noisy scans. Use low to medium strength. Apply via a masked layer so you only affect the face, not the whole scene.
• Watch for identity drift (eyes too glossy, altered expressions). If a face looks “off,” reduce strength or revert.

9) Upscaling for print or display

• Target output: For prints, 300 ppi at final size is a safe baseline; posters can go lower depending on viewing distance.
• Use an AI upscaler (Real‑ESRGAN, Topaz Gigapixel) after cleanup, not before. Choose models that preserve texture without inventing heavy micro‑detail.
• Compare 1× vs 2× vs 4×; bigger isn’t always better. Stop when added detail looks synthetic.

10) Finishing and export

• Add gentle grain if denoise/upscale made the image too plasticky; it can restore a natural look.
• Export a high‑quality master TIFF (16‑bit), and separate JPEG for sharing. Embed color profiles.
• Keep layered PSD/PSB or .kra/.xcf for future edits.

Practical Workflows for Common Scenarios

Below are concrete end‑to‑end workflows you can replicate. Adjust steps to your toolset.

Workflow A: 1940s sepia family portrait with cracks

1. Scan at 600–1200 dpi, 16‑bit TIFF. Keep sepia tone for the first pass.
2. Dust & Scratches filter masked into the background; manual healing on clothing.
3. Inpainting with Generative Fill for background cracks: prompt “continue wall texture; no new objects.” Keep opacity ~70% to blend.
4. Curves to boost midtone contrast; separate layer for delicate dodge on faces.
5. Optional: Convert to monochrome + add subtle split toning for sepia recreation.
6. Gentle face restoration (CodeFormer weight low, e.g., 0.3) masked to cheeks and eyes only.
7. Upscale 2× with Real‑ESRGAN; add a touch of grain.
8. Save master TIFF and a social JPEG.

Workflow B: 1998 indoor snapshot, heavy noise and color cast

1. Camera capture of print with diffuse window light, deskew, crop.
2. White Balance using a neutral wall; HSL to reduce magenta cast.
3. AI denoise (Lightroom Denoise at a conservative strength). Minimal sharpening.
4. Tone with Curves; recover shadows softly to avoid banding.
5. Optional: Remove red‑eye; heal small dust.
6. Upscale 1.5–2× for modern displays; export.

Workflow C: 1970s color print faded yellow

1. Flatbed scan 600 dpi, 16‑bit.
2. Levels/Curves per channel to neutralize cast; use a gray card reference if available (even a newspaper edge in the scene can help).
3. HSL to reduce oversaturated reds; local adjustment brush for skin.
4. Optional: Use a LUT that approximates the original film stock if known; blend at low opacity.
5. Light de-noise, add gentle texture; export.

Workflow D: 35mm negative scan with dust and scratches

1. If your scanner supports Digital ICE (infrared channel), enable it; it automatically maps dust.
2. Manual heal remaining marks; avoid smearing grain.
3. Use a film‑friendly denoise (grain-aware) to avoid waxy look.
4. Curve and color balance, then optional Real‑ESRGAN with a model tuned for film grain.

Tool-by-Tool: Practical Settings That Work

• Topaz Photo AI
  • Start with Auto, then dial back Remove Noise and Sharpen to avoid plastic results. Use Face Recovery at 0.2–0.4. Compare before/after at 100%.
• Photoshop
  • Spot Healing Brush on “Content-Aware,” Sample All Layers. For Generative Fill, keep prompts literal: “Continue background wood texture,” “Recreate sleeve fabric.” Save variations and blend.
• Real‑ESRGAN
  • Use realesrgan-x4plus for general prints, x4plus-anime is not suitable for photos. Consider realesrgan-x2plus if artifacts appear at 4×.
• GFPGAN/CodeFormer
  • Apply via a separate layer and mask. Start with low strength; overuse creates identity shifts.
• Stable Diffusion Inpainting (local)
  • Use a model fine‑tuned for inpainting. Set denoise around 0.2–0.35 for small repairs. Enable ControlNet (Canny/Lineart) to preserve structure. Keep CFG moderate.
• GIMP/Krita
  • Heal vs Clone: Heal blends texture; Clone copies exactly. For paper texture areas, Heal is often better. Work on separate layers.

A Recovery Checklist You Can Reuse

Use this checklist to avoid missing critical steps.

• Before you start

  • Digitize at 600+ dpi (prints), 16‑bit TIFF if available
  • Clean the print gently; stabilize lighting for camera captures
  • Set up a dated folder structure; duplicate the original file read‑only
  • Decide preservation vs presentation vs reconstruction

• During restoration

  • Deskew and crop with margin
  • Remove dust and specks first; inpaint tears with tight masks
  • Denoise only as much as needed; avoid halos
  • Tone via Curves; correct color cast per channel
  • Optional: colorize with documentation; face restore lightly and locally
  • Upscale after cleanup; compare scales for realism

• Finishing and archiving

  • Add subtle grain if needed; soft proof for print
  • Export 16‑bit TIFF master + layered working file + JPEG for sharing
  • Embed ICC profiles; write basic provenance/notes in metadata
  • Back up to at least two locations (local + cloud or drive)

Common Mistakes That Ruin Restorations

• Over‑smoothing and plastic skin: AI denoise at high strength removes natural texture. Dial back and reintroduce subtle grain.
• Identity drift in faces: Aggressive face restoration can alter expressions and features. Always mask and use low strength.
• Incorrect order of operations: Upscaling first can magnify defects. Clean before you enlarge.
• Clipped channels during color correction: Watch histograms per channel; avoid pushing Curves into hard clipping.
• Repeated JPEG saves: Each save degrades quality. Keep a TIFF/PSD master and export JPEG only once.
• Auto white balance on sepia/aged prints: It can remove the intended tone. Preserve or recreate the tone intentionally.
• Over‑creative inpainting: Generative tools can insert objects or patterns that didn’t exist. Keep prompts factual and masks tight.
• Cropping too tight: Future framing or edge context may be lost. Preserve borders unless damaged.

Ethics, Attribution, and Provenance

• Label reconstructions: If you colorize or rebuild missing areas, note it in the filename or metadata (e.g., Reconstructed_2026_by_[Name]).
• Preserve the original: Keep the raw scan and a non-destructive working file.
• Respect privacy: Family photos often contain sensitive content. Prefer local processing; read cloud terms.
• Cultural context: Avoid imposing modern aesthetics (e.g., extreme skin smoothing) on historical images.

Quality Control: How to Know You’re Done

• View at 100% and at typical display sizes. If detail looks “too good to be true” at 100% but natural at fit-to-screen, consider dialing back.
• Print a small proof (4x6). Physical prints reveal banding, casts, and oversharpening that screens hide.
• Get a second set of eyes. Others may catch uncanny faces or repeated texture patterns.
• Keep versions: v1_cleanup, v2_inpaint, v3_tone, etc. You should be able to roll back any step.

Mini Decision Frameworks for Tricky Calls

• Denoise or not?
  • If grain is uniform and not distracting, keep it. Grain carries texture and history. Denoise only to remove color blotches and scanning noise.
• Colorize a B&W?
  • If the goal is historical presentation, keep monochrome. For wider sharing, colorize but publish both.
• Upscale 4× or 2×?
  • Compare hairlines, fabric weave, and edges. If they look painted at 4×, step back to 2×.
• Which face restorer?
  • GFPGAN is stronger; CodeFormer is often subtler. Start with CodeFormer at low strength.

FAQs

• What resolution do I need for printing restored photos?

  • Aim for 300 ppi at the target print size. A 2400×3000 px image prints well at 8×10 inches. Larger wall prints can go lower (200–240 ppi) due to viewing distance.

• Can I restore photos taken on a phone of a print?

  • Yes. Use diffuse light, avoid glare, and deskew. Quality may be lower than a flatbed scan but can be sufficient for social sharing and small prints.

• Should I colorize old black‑and‑white photos?

  • It’s a creative choice. Keep a monochrome master and label colorized versions as interpretations.

• How do I handle water‑damaged or moldy prints?

  • Digitize first to prevent further handling damage. Inpainting and careful tone recovery can help. If the print is physically wet or moldy, consult conservation guidance before cleaning.

• Are online AI services safe for sensitive family photos?

  • Read terms carefully. Some services retain images for model training or analytics. For sensitive images, process locally.

• What file format should I use to save my master?

  • 16‑bit TIFF with embedded ICC profile. Keep layered PSD/PSB (Photoshop) or .kra/.xcf (Krita/GIMP) for non-destructive edits.

• Can AI bring back information that isn’t there?

  • AI can hallucinate plausible detail but not true historical data. Treat reconstructions as best-effort interpretations.

• How do I restore negatives and slides?

  • Use a dedicated film scanner or a camera + macro lens with a light panel. Enable infrared dust removal if available. Treat color casts per channel.

• What about color banding after edits?

  • Work in 16‑bit where possible, avoid extreme local contrast, and add subtle grain to mask banding.

Put This Into Practice With an AI Agent

An AI agent can orchestrate your restoration pipeline, keep your process consistent, and batch work safely. Here’s a practical blueprint you can implement in a workspace like Vife Agent.

Agent goals

• Intake and classify images by damage type (dust, scratches, tears, color cast, noise).
• Recommend a step-by-step plan per image, tailored to your privacy and toolset (local vs cloud).
• Generate prompts/settings for inpainting, denoise, face restore, and upscaling.
• Track versions and export masters with embedded metadata.

Inputs and structure

• Inputs: Folder of scans, a short survey (Preservation vs Presentation vs Reconstruction), privacy preference, target output size.
• Tools: Your chosen stack (e.g., Photoshop/Lightroom + local SD Inpaint + Real‑ESRGAN + CodeFormer), or a desktop all-in-one.

Example agent workflow

1. Ingest
   • Scan folder, read EXIF/ICC, and flag low-resolution or JPEG-only sources.
2. Assess
   • Classify images (e.g., “tear present,” “heavy noise,” “yellow cast”). Create a to-do per asset.
3. Plan
   • Propose a pipeline with specific tool settings. Example plan for “small tear + mild noise + yellow cast”:
     • Heal dust manually; Generative Fill tear repair with prompt: “Continue background paper texture; no new objects.”
     • Lightroom Denoise at 15–25; Curves per channel to neutralize cast.
     • Optional CodeFormer at 0.2 masked to the face; Real‑ESRGAN 2×.
4. Execute (semi-automated)
   • Launch scripts or actions where possible (e.g., Photoshop Actions for Dust & Scratches pre-mask, batch Real‑ESRGAN upscaling). Pause for human review after each major step.
5. QC
   • Prompt the user: “At 100% zoom, check for plastic skin and repeated patterns in background. Approve or request revision.”
6. Export and archive
   • Save TIFF master + layered file + JPEG, embed metadata note: “Restored YYYY‑MM‑DD; inpainting on background only; no face alteration.”

Prompt and setting templates

• Inpainting (Photoshop Generative Fill)

• Stable Diffusion Inpaint (local)

• Real-ESRGAN

• CodeFormer

Automation tips

• Batch rename and version with a naming schema: YYYYMMDD_Project_ImageID_vN.tif.
• Use watch folders to automatically run upscaling after QC approval.
• Store agent decisions in a sidecar JSON per image for provenance.

Advanced Topics (When You Need More)

• Paper texture simulation: To blend repaired areas, sample paper texture with a high‑pass layer and overlay at low opacity.
• Channel‑wise retouching: Work on individual RGB channels to remove scratches that only appear in one channel.
• Frequency separation for prints: Use low‑frequency layer for tone and high‑frequency for texture; heal on the appropriate layer to avoid smearing.
• Recreating borders: If edge borders are damaged, generate a new border on a separate layer, keeping the original inside intact.
• Negative space completion: Where a corner is missing, clone from nearby and use inpainting to randomize repetition.

From “AI Magic” to Repeatable Craft

The difference between a quick fix and a keeper is process discipline. By capturing well, restoring in a sensible order, and applying AI conservatively, you’ll produce images that feel authentic—not uncanny.

Conclusion: Restore, Preserve, Share

AI image restoration is best when it respects the original and saves your time. Start with careful digitization, follow an ordered workflow, choose tools based on privacy and control, and apply AI gently. Document what you change, keep lossless masters, and print a proof before you commit to large runs.

If you want to keep momentum, set up a small restoration agent to draft steps, generate prompts, and batch routine tasks. You can continue this work inside Vife Agent—import this checklist, wire up your tools, and let the agent keep you fast and consistent while you focus on judgment and taste.