how to remove scratches from old photos without losing the original image?

What "Scratch Removal" Actually Means for an Old Photograph

When people search for how to remove scratches from old photos, they usually picture a single crack across a face in a 1940s portrait. In practice, an old physical print carries several layered defects at once: silver mirroring, silver oxidation, gelatin emulsion shrinkage, dust, mold, fading, and physical scratches. A "scratch" is specifically a linear defect that exposes the paper base beneath the emulsion, while a "crack" follows the gelatin layer. Digital photograph restoration treats both, plus dust, dirt, and aging marks, by using pixel-level healing tools in a raster graphics editor or by feeding the image into a neural network trained on paired damaged/clean examples. Knowing the difference matters because scratch-specific tools (spot healing, content-aware fill, inpainting) behave differently on each defect, and the wrong tool creates halos, color drift, or ghosting around faces.

Also worth reading: What is the best AI video enhancer in 2026 for colorizing and upscaling footage without losing quality? · Can AI really restore old family photos without ruining them, and what should you know before trying it? · How can I ensure AI colorization of historical photos remains accurate to the era without hallucinating details?

For this reason, scratch removal should always start with the highest-resolution scan you can produce (600 DPI minimum, 2400 DPI for 35mm), saved as a 16-bit TIFF or lossless PNG. The master file stays untouched on a backup drive, and every edit happens on a duplicate. That workflow protects the original artifact — the actual print — from repeated handling, because the original can continue to degrade even while you work on a digital copy. A 2024 PetaPixel review of ON1 Restore AI reported that even aggressive neural "auto-fix" passes produced grotesque results on faces with deep creases, reinforcing that a scan-only, non-destructive workflow is the safer starting position.

The Two Real Methods: Manual Raster Editing vs. AI Inpainting

The first method is manual restoration in software like Adobe Photoshop, GIMP, Affinity Photo, or Paint.NET. You clone pixels from a clean neighboring area over the scratch, use the Spot Healing Brush for small dust spots, and use Content-Aware Fill or the Healing Brush for long lines. This gives full control but demands skill: a 2022 root-nation.com roundup of five restoration programs noted that manual tools require 2–6 hours of practice before a beginner can hide a single hairline scratch convincingly. The output quality is high once learned, but the time cost per photo is real.

The second method is AI-based restoration. Modern tools such as Remini, Adobe Firefly's Generative Fill, the Google Nano Banana Pro model, and ChatGPT's image-repair prompts work by detecting scratches as missing pixels and hallucinating plausible texture based on context. Reviews in 2025–2026 from perfectcorp.com and criticalhit.net both reported that leading consumer apps correctly handled 80–90% of light surface scratches but consistently failed on heavy emulsion cracks and on textured backgrounds (brick walls, foliage, patterned wallpaper) where the model invents detail that was never there. AI is dramatically faster — most apps return a restored image in under 60 seconds — but it is not a substitute for a careful human pass on irreplaceable family portraits.

A reasonable middle path is what restoration studios call the "AI pass + human pass" hybrid: run an AI cleanup first to remove dust and light scratches in bulk, then open the result in Photoshop and use the Clone Stamp to correct the residual errors, especially around eyes, jewelry, and printed text.

Step-by-Step: A Practical Workflow That Actually Works

Start by inspecting the physical print under raking light (a lamp at a low angle) to identify whether damage is surface dust, silver-mirror fade, or actual scratches through the emulsion. Use a soft brush or canned air to remove loose dust; never touch the surface with bare fingers because skin oils accelerate emulsion decay. Scan at 600 DPI for 4×6 prints and 1200–2400 DPI for smaller originals like wallet-size or film negatives, in 16-bit color to preserve tonal range. Save as TIFF; PNG is acceptable but produces larger files for the same quality.

Open the scan in your editor of choice. Convert to a working profile (ProPhoto RGB or Adobe RGB) so you have headroom for tonal adjustments later. Duplicate the layer and work non-destructively — every heal, clone, or AI fill should sit on its own layer with a mask so you can undo selectively. For dust and small spots, the Spot Healing Brush with "Content-Aware" mode is usually sufficient. For a single long scratch across a cheek or forehead, sample a clean patch with the Clone Stamp at low opacity (30–50%), build up multiple passes, and feather the brush edge to 0% hardness. Run a final pass of the Camera Raw or Lightroom noise-reduction filter at low strength (10–20) to smooth any banding the cloning introduced.

If using an AI tool, upload the same high-res scan, run a single restoration pass, then compare side-by-side with the original at 100% zoom. If the AI has softened facial pores, dropped eyelashes, or hallucinated buttons on a jacket, undo and rely on manual tools for the affected zone. Save the AI result as a separate file from the manual edits; never overwrite the master scan.

Tool Comparison: Manual Editors, AI Apps, and Hybrid Pipelines

Below is a side-by-side comparison of the most common categories of tools available as of September 2026, based on the 2026 perfectcorp.com hands-on review and the criticalhit.net 2026 editor showdown.

The hybrid column is the realistic best option for a high-value family photograph. Pure AI is appropriate when you are cleaning hundreds of snapshots where the loss of a few is acceptable; pure manual editing is appropriate for one image you cannot re-scan.

Common Mistakes That Permanently Damage the Image

The most common error is scanning at too low a resolution. A 300 DPI scan of a 4×6 print gives roughly 1200×1800 pixels — enough for a phone screen but not enough to heal a 1mm scratch convincingly. The minimum useful scan for restoration is 600 DPI, and 1200 DPI is preferable. Another mistake is running an AI face-smoothing pass before scratch removal; the smoother surface gives the AI less texture to work with, and the inpainting then invents detail rather than reconstructing it. Always remove scratches first, then optionally apply portrait enhancement.

A third mistake is using the JPEG format as the working file. JPEG compresses in 8×8 blocks, and every save discards data. A file saved as JPEG ten times has visibly more banding around high-contrast edges than the same file saved as TIFF or PNG even once. Always work in lossless formats and only export to JPEG for sharing. A fifth frequent error is "over-healing": cloning aggressively until the photo looks like an airbrushed portrait. Restoration aims to recover what the photo once was, not to produce an idealized version, and forensic or historical prints should retain their grain.

Finally, never trust an AI tool's "auto-enhance" toggle. The PetaPixel review of ON1 Restore AI documented cases where the algorithm rebuilt missing facial features incorrectly, and similar problems have been reported in test reviews of consumer apps. Verify every change at 100% zoom before saving.

When to Act, When to Hire a Specialist

Act now if your photos show early signs of silver mirroring (a metallic sheen on dark areas) or if the prints are stored in a humid environment. Once silver mirroring progresses, it becomes chemically irreversible — no software can restore the metallic particles. Hire a professional conservator (the American Institute for Conservation maintains a directory) if the original is a 19th-century tintype, daguerreotype, or ambrotype, where emulsion chemistry differs from 20th-century gelatin prints. For paper damage, separating a photo that has fused to its glassine sleeve, or for negatives stored on unstable nitrate film, professional intervention is the only safe path.

If the photo is a mid-20th-century gelatin print with moderate scratches and fading, the manual-plus-AI workflow above is achievable in a single afternoon. The cost of doing it yourself is your time and a Photoshop subscription (around $23/month as of 2026) or a free tool like GIMP. The cost of a professional digital restoration is typically $25–$75 per image at a standard service and $150+ per image for a museum-quality archivist, with most studios requiring a 600 DPI scan supplied by the client.

Pricing Reality Check for AI Restoration Tools in 2026

The consumer market in 2026 has settled into three pricing tiers. The free tier — Remini's daily quota, PhotoMyne's basic mode, and the default ChatGPT image-repair capability — produces usable but inconsistent results and is best for low-stakes batches. The subscription tier ($5–$15/month per app) usually removes watermarks, raises resolution caps, and unlocks batch processing; Adobe's Photography plan at roughly $23/month bundles Photoshop, Lightroom, and Firefly under one license. The professional tier combines Photoshop with a dedicated AI plugin such as ON1 Restore AI, Topaz Photo AI, or the Imagen AI retoucher plugin, and typically runs $100–$300/year total.

A reasonable budget for restoring a single family album of 200 prints is therefore either your time (DIY at zero software cost with GIMP) or $25–$100 if you send scans to a restoration service. AI tools are not a substitute for the service when the originals are historically or personally irreplaceable.

What AI Colorization Has to Do With Scratch Removal

Although the question is about scratches, colorization and scratch removal are usually performed together. Modern colorization tools — including colorizethis.io, MyHeritage's In Color, and the open-source DeOldify project — assume a clean, scratch-free input. Feeding them a heavily scratched black-and-white photo produces a colored image where every scratch is colored too, drawing attention to the defect instead of hiding it. The accepted workflow is therefore: scan → restore scratches → colorize → optional final retouch. AI colorization models released since 2023 are trained on large color-photo corpora and can hallucinate plausible skin tones, sky tones, and fabric colors, but they cannot infer original surface texture, so scratch removal must come first.

This is also why photo-restoration tutorials now bundle the two operations into one prompt or one pipeline. Recent guides from Blockchain Council, perfectcorp.com, and The Indian Express (covering Google Nano Banana Pro) all describe a two-pass sequence in which an "old-photo restoration" prompt is followed by a "realistic colorization" prompt. The sequence works precisely because scratch removal and colorization are separable, ordered operations — colorization is the final interpretive step, not a substitute for repair.

Final Recommendations

For a single treasured family photograph, scan at 1200 DPI in 16-bit TIFF, then use Photoshop's Spot Healing Brush and Clone Stamp on a duplicated layer; expect 30–90 minutes of careful work. For a batch of dozens of casual snapshots, run them through a reputable AI app such as Remini or PhotoMyne, then inspect the top 10% at full resolution and manually correct any face-level errors. For daguerreotypes, tintypes, or visibly chemically unstable prints, consult a paper conservator rather than a software tool. Save the original scan forever, never overwrite it, and treat every AI suggestion as a draft rather than a finished restoration.