creative

Best AI Photo Restoration Tools

The best AI photo restoration tools in 2026 — Google Gemini with Nano Banana 2 and Nano Banana Pro, Adobe Photoshop, Remini, MyHeritage, VanceAI, Topaz Photo, Fotor, Hotpot.ai, Palette.fm and the open-weight models — compared on damage repair, face reconstruction, colourisation, published pricing, licence restrictions and what the independent benchmark evidence actually says about identity accuracy.

Updated September 2026

Quick answer: The best AI photo restoration tool in September 2026 for a damaged print — cracks, tears, stains, missing corners — is Google Gemini, because its image models rebuild the frame coherently rather than patching the damaged region. Free Gemini accounts run Nano Banana 2, which has been the default image model across the Gemini app, Google AI Mode and Google Lens since 26 February 2026; Nano Banana Pro is the stronger model and stays with paid Google AI Pro and Ultra plans, reached by regenerating an image from the three-dot menu (Google; Google DeepMind). The best tool for a face that has gone soft or grainy is Remini, which caps output at 2x and 4096 x 4096 pixels (Remini). The best route for colourising a family archive is MyHeritage, which is the only major platform selling a photo-only subscription alongside its genealogy plans (MyHeritage). The best offline option is the open-weight stack — but check the licence, because CodeFormer ships under NTU S-Lab License 1.0 and cannot be used commercially, while Microsoft’s Bringing Old Photos Back to Life is MIT-licensed and unrestricted. The caveat that governs the whole category: every one of these tools invents pixels. In the only independent 2026 benchmark of real-world face restoration, the organisers ran a separate identity checker alongside image-quality scoring precisely because the best-looking output is frequently not the same person (NTIRE 2026, arXiv:2604.10532).

Prices, model versions and feature claims were checked on 17 September 2026. Several vendors in this category run region-dependent or device-dependent pricing, and several publish figures that third-party trackers contradict. Where sources disagree, the disagreement is shown rather than averaged, and figures we could not stand behind are marked “data not available”.

Restoration is not upscaling, and four different jobs hide behind one search term

The single most expensive mistake in this category is buying an upscaler when you needed a restorer. They are different jobs with different failure modes.

Upscaling adds pixels. It takes a small or soft file and produces a larger one. It does not know that the diagonal white line across your grandmother’s face is a crack in the emulsion, and most upscalers will happily sharpen that crack into a crisper crack. If that is your problem, read our best AI image upscalers guide instead.

Restoration repairs damage and recovers appearance. Four distinct jobs sit underneath it, and no single tool is best at all four:

1. Damage repair. Cracks, tears, creases, tape marks, mould, foxing, water stains, chunks missing from a corner. This is an inpainting problem: the tool has to invent content that was never in the file. Generative models are now clearly the strongest option here, and traditional editors remain the best option when you need to control exactly what gets invented.

2. Face reconstruction. A face that has gone soft, grainy or low-resolution. This is the most studied problem in the category, it has its own academic name — blind face restoration — and it carries the highest risk, because the failure mode is a plausible stranger rather than an obvious artefact.

3. Colourisation. Turning a black-and-white or sepia image into colour. This is a guess by construction, and the research is unambiguous about that. It has its own section below.

4. Tonal recovery. Fading, colour casts on 1970s and 1980s prints, yellowing, low contrast. This is the least glamorous job and the one most often solved for free by tools you already own.

A vendor that sells more than one of these is ranked here on its restoration product line only. Topaz sells Gigapixel for enlargement and Topaz Photo for repair, and only Topaz Photo is in scope on this page. Remini enlarges as well as restores, and is ranked here on face reconstruction rather than on its 2x ceiling. For tools that edit an undamaged photograph, see best AI photo editor; for creating an image from a prompt, see best AI image generator.

The state of AI photo restoration: September 2026

Four things have changed the category in the last twelve months, and none of the affiliate roundups currently ranking for this term mention more than one.

General-purpose image models have overtaken dedicated restoration tools. Google announced Nano Banana Pro, its Gemini 3 Pro Image model, on 20 November 2025, and launched Nano Banana 2 — built on Gemini 3.1 Flash Image — on 26 February 2026, making it the default image model across the Gemini app, Google AI Mode and Google Lens (Google). Both models reached general availability for enterprise on 28 May 2026 (Google Cloud). What matters for restoration is the architecture: these models construct interim intermediate representations before producing the final frame, which is why they repair a crack across a face by regenerating the region coherently rather than cloning nearby texture into the gap. A dedicated restorer built on a 2021-era GAN cannot do that at any price. Which model you get depends on what you pay: free Gemini accounts run Nano Banana 2 with a reported daily image allowance of roughly 20 images, while Google says Pro and Ultra subscribers keep access to Nano Banana Pro for specialised tasks by regenerating an image from the three-dot menu (Google).

Photoshop’s dedicated restoration filter has been unreliable. Adobe’s Photo Restoration Neural Filter returned “We’ve temporarily disabled this filter because of an error” on Photoshop versions 26.11.2 and 27.1.0, in a bug thread on Adobe’s own community forums opened in December 2025. Adobe stated that the 27.3 beta included Neural Filter fixes; an Adobe community manager then confirmed on 15 January 2026 that the issue still reproduced in 27.2 and 27.3 beta and escalated it, and the last user reply on 23 January 2026 reported the beta had not solved it. The thread was never marked resolved and we found no evidence of a fix since, so the filter’s current status is data not available. It is named as the restoration method in a large share of the guides ranking for this keyword, so check that it runs on your install before you plan a workflow around it. The Colorize Neural Filter is a separate filter and was still listed as available.

The best restoration tool now watermarks your family photograph. Google made the visible sparkle mark on Gemini output optional in August 2026, but the invisible SynthID signal and the C2PA metadata stay (TechCrunch, 14 August 2026). Separately, Google Photos adds Content Credentials to a file when an AI edit is saved and exposes them under a “How this was made” section that splits provenance into media composition, AI edits, possible AI edits and non-AI edits (Google Photos Help). This is covered in full below, because for a photograph destined for a family archive it is a feature, and for a photograph destined for a newspaper or a court it is a problem.

The archive profession has published guidance, and it does not say “use AI”. The Archives and Records Association (UK and Ireland) issued AI Preparedness Guidelines for Archivists in February 2026, authored by Professor Giovanni Colavizza and Professor Lise Jaillant as part of the FLAME project (Towns Web Archiving). The guidance is about governance and documentation rather than tool adoption, and the wider professional consensus in 2026 is that a colourised or reconstructed image must be labelled as such so the distinction between the original document and the modern interpretation survives.

Best AI photo restoration tools ranked (2026)

Ranking criteria and the limits of this ranking are set out under How we rank.

The tools in depth

1. Google Gemini — best overall for damaged prints

Type: General-purpose image models, prompt-driven Price: Free tier runs Nano Banana 2; Google AI Pro at $19.99 a month keeps access to Nano Banana Pro, regenerated from the three-dot menu, with higher quotas on the Ultra tier Strongest at: Cracks, tears, missing regions, creases, mixed damage Weakest at: Repeatability, and anything that has to be evidentially faithful

Nano Banana Pro is built on Gemini 3 Pro and is Google’s most capable image editing model, offering identity preservation across multiple subjects in one frame, localised edits, and 2K and 4K output (Google DeepMind). Nano Banana 2, the free default since 26 February 2026, is built on Gemini 3.1 Flash Image and is materially faster and cheaper to run, which is why Google made it the default rather than Pro.

Why it wins: it treats a damaged photograph as a scene to be re-rendered coherently rather than a hole to be filled. A crack that crosses a face, a shoulder and a background wall is the case where every patch-based tool fails and this one does not. The free tier is also the best free restoration available, which no dedicated restoration product on this list matches.

Limitations: output is not deterministic — the same prompt on the same scan gives different results, which makes a consistent pass over 200 photographs hard. Every output carries SynthID and C2PA provenance data. Free-tier daily image allowances are reported at roughly 20 images and are not contractually published, so plan a large archive around a paid tier. And because the model regenerates rather than repairs, it will confidently invent a plausible ear, collar or background that was never in the original.

Best for: one valued print with real physical damage, where you will look at the result carefully before accepting it.

2. Adobe Photoshop — best control over what the AI invents

Type: Professional editor with generative tools Price: $22.99 a month for Photoshop as a single app; the Creative Cloud Photography plan is $19.99 a month with 1TB. The 20GB Photography plan at $14.99 a month is not offered to new subscribers — it does not appear on Adobe’s current plans page and now exists only as a grandfathered rate. Strongest at: Selective, reversible, documented repair Weakest at: Speed, and its own dedicated restoration filter

Photoshop is the only tool here where you decide exactly which pixels the model is allowed to touch. Generative Fill on a selected tear, the Remove tool on tape marks, and a manual layer stack give you a repair you can explain, undo and hand to someone else.

Why it ranks second rather than first: its purpose-built restoration feature has been unreliable. The Photo Restoration Neural Filter was reported disabled with an error on Photoshop 26.11.2 and 27.1.0 in Adobe’s community bug threads from December 2025, confirmed by Adobe as still reproducing in January 2026, and never marked resolved. The generative tools that do the actual restoration work are unaffected.

Limitations: it is a subscription, it has a real learning curve, and restoring one photograph properly in Photoshop takes longer than the entire Gemini workflow.

Best for: a photograph that matters enough to repair deliberately, or any job where you must be able to say what was changed.

3. Remini — best face reconstruction, worst pricing transparency

Type: Mobile-first restorative enhancer Price: Data not available. Remini publishes no price list and runs region-dependent and device-dependent price tests. Third-party trackers this run reported roughly $6.99 a week for a personal plan, roughly $9.99 a week for an unlimited plan, and around $9.99 a month or about $50 a year for Pro — figures which do not reconcile with each other and should be treated as unverified. Maximum output: Remini states 2x enlargement and resolution “up to 4096x4096px” (Remini), without distinguishing tiers; free exports are reported to be capped lower, at around 2080 x 2080 pixels Owner: AI Creativity S.r.l., a Bending Spoons company

Remini’s specialism is reconstructing faces from low-quality source material, and its mobile flow is the easiest here for a non-technical user. Practitioner consensus puts it ahead of other consumer apps on this specific job, though as the accuracy section below explains, no independent 2026 test scores consumer restoration apps against each other on identity fidelity, so treat that consensus as consensus rather than measurement.

Limitations: the free tier is ad-gated and watermarks output. The 2x, 4096-pixel ceiling means it is not an enlargement tool despite being sold alongside them. And the reconstruction is aggressive: Remini produces a sharp, modern, slightly smoothed face, which is what most people want and what an archivist would object to.

Best for: soft or grainy faces in a phone-sized archive, where you accept a flattering reconstruction.

4. MyHeritage — best for colourising a family archive

Type: Genealogy platform with photo tools Price: Conflicting, and worth checking carefully. MyHeritage’s own help centre lists six subscription types — Premium, PremiumPlus, Data, Complete, Omni and Photo (MyHeritage). The Photo plan covers unlimited scanning, colourisation, enhancement and damage repair without the genealogy features, and MyHeritage does not publish its price in a form we could verify this run. Third-party trackers give irreconcilable figures for the genealogy plans — one set reports $119, $189 and $259 a year, another reports Premium at $89 first year rising to $129 on renewal and Complete at $199 rising to $299 — so do not treat any single figure here as final. The help centre lists no “Basic” paid tier, so treat comparison pages that show one with caution. Free allowance: 10 images, watermarked Strongest at: Colourisation, enhancement and keeping results attached to a family tree Weakest at: Published pricing clarity

MyHeritage’s In Color, Photo Enhancer, Photo Repair and Reimagine tools are the most-used colourisation tools in genealogy for a reason that has nothing to do with model quality: the restored image lands next to the person it depicts, in a tree, with a source citation. That context is the product, and the existence of a photo-only subscription means you no longer have to buy a genealogy plan to get it.

Limitations: the pricing is the weakest part of the proposition, and sources disagree on which tier unlocks unlimited colourisation, with at least one stating the top Complete tier is required. Free output is watermarked.

Best for: anyone doing family history, and anyone who wants colourisation kept alongside the people it depicts.

5. Topaz Photo — best local processing, with an ownership question attached

Type: Desktop application, runs locally Price: $39 a month or $199 a year for Personal, and $599 a year for Pro, on Topaz’s own pricing page as read on 17 September 2026. Only subscriptions are sold; perpetual licences are no longer offered. Strongest at: Recovery and face recovery models on your own machine, no upload Weakest at: Cost, and corporate certainty

Topaz Photo — the product Adobe’s acquisition release calls Topaz Photo, formerly marketed as Topaz Photo AI — has the strongest restoration available without sending the file anywhere. For sensitive images, including medical, legal or anything you will not upload, this is effectively the only serious commercial option.

The ownership question: Adobe agreed to acquire Topaz Labs on 25 June 2026, with the transaction expected to close in the second half of 2026 subject to regulatory approval. Our check on 17 September 2026 found no announcement that the deal had closed. Coverage at announcement said the apps continue as standalone products and existing perpetual licences stay valid, but a subscription bought now is a subscription to a company mid-acquisition.

Best for: offline work, batch processing on your own hardware, and anyone who will not upload the file.

6. VanceAI Photo Restorer — batch work on credits, with pricing you must check

Type: Web platform, credit-based Price: Conflicting across every source we checked. VanceAI’s own pricing page was reported this run as showing four monthly plans at $9 for 200 credits, $17 for 500, $26 for 1,000 and $42 for 2,000 on a 50 per cent annual discount, with a 10-credit free trial. Third-party trackers instead report subscriptions at $9.90 to $9.95 a month for 100 credits rising to $59.95 for 1,000, plus one-off credit packs at $4.95 for 100, $7.95 for 200, $12.95 for 500 and $17.95 for 1,000. Credit consumption is disputed too: third parties say one to three credits per image, while VanceAI’s own tool table was reported as charging four credits per restoration up to 4K and eight credits at 4K. On the higher figure, 100 restorations cost roughly 400 credits, not 100. Strongest at: Running many photographs through the same treatment Weakest at: Knowing what it will cost

VanceAI is a genuine batch tool and the interface is built for volume. We are not quoting a headline per-photo price for it, because the two variables that determine that price — plan cost and credits per restoration — both have conflicting published values, and multiplying two disputed numbers produces a third number nobody should rely on.

Limitations: rollover is conditional. Subscription credits roll over while you remain subscribed and expire when the subscription ends, with rollover reported as capped at five times the monthly allowance — so this is not a way to bank credits and cancel.

Best for: a large archive, once you have opened the pricing page and confirmed both numbers yourself.

7. Hotpot.ai — cheapest per-photo with no subscription

Type: Web tool, pay per image Price: Reported free tier of two to three uses per tool, then pay-per-photo from about $0.10 Strongest at: Scratch and blemish removal on a handful of images Weakest at: Volume, and faces

Hotpot’s AI Picture Restorer removes scratches and imperfections and sharpens faces. The reason it is on this list is the pricing model: roughly ten cents an image with nothing recurring is the cheapest straightforward entry point in the category.

Limitations: the free tier is very small, and results do not match a generative model on serious damage.

Best for: three or four photographs, paid for once.

8. Palette.fm — best dedicated colouriser

Type: Web colourisation tool with prompt control Price: Conflicting. Third-party sources this run reported a free preview tier, annual subscriptions from about $6 a month, an annual plan around $72, per-image costs of $0.10 on the basic plan and $0.05 on Pro, and credits at about $0.01 each. These do not reconcile into a single coherent price list. Free output is reported at 500 x 500 pixels with a watermark; paid output up to 5000 x 5000 pixels without one. Strongest at: Steering colour choices instead of accepting the model’s first guess Weakest at: Cost at archive scale, and price transparency

Palette.fm’s advantage over a one-click colouriser is that you can direct it with a prompt — which matters, because as the next section explains, colourisation is a guess and the person who knew the real colours is usually the reader, not the model.

Limitations: credit-based pricing on a large family archive adds up quickly, and the published pricing is among the least consistent of any tool on this page.

Best for: a small number of black-and-white photographs where you know something about the true colours.

9. Fotor — quick browser restoration, but not the free option people claim

Type: Web editor with an AI restoration tool Price: Reported between $3.33 a month on annual billing and $8.99 a month depending on source and plan — another conflict we could not resolve Free tier: Reported as 6 starter credits on a free account, with every free export watermarked and the AI photo enhancer capped at roughly one use a day Strongest at: A usable result in under a minute with nothing installed Weakest at: Heavy damage, and the free claims made on its behalf

Fotor is widely listed as the best free photo restorer. That claim does not survive checking: independent reviews this run report that a free account is required, that starter credits are finite, and that free exports carry a watermark. Vendor marketing offering complimentary watermark-free credits to new users still describes an account with a finite balance, not an unlimited free tool.

Limitations: it is a general editor with a restoration feature, not a restoration specialist. On a torn print it smooths rather than rebuilds.

Best for: a lightly faded photograph when you want a one-click result and do not mind signing up.

10. The open-weight stack — free, unlimited, and the licences differ more than people think

Three open models do most of the work in this category, including inside several commercial products. Their licences are not the same, and this is the most commonly misstated fact in the category.

Bringing Old Photos Back to Life (Microsoft, arXiv:2004.09484) is MIT-licensed, covering both the code and the pretrained model, which makes it the only genuinely unrestricted option here. It remains the reference implementation for structured physical damage — scratches, tears and fading — rather than face quality, and is the right pipeline when the problem is the print rather than the person.

GFPGAN (Tencent ARC Lab) uses a StyleGAN2 generative prior and states that GFPGAN is released under Apache License Version 2.0. It is not restricted the way CodeFormer is, but it builds on third-party projects and pretrained components that carry their own terms, so check the dependencies your pipeline actually ships before using it commercially.

CodeFormer uses a codebook lookup with an adjustable fidelity weight. It is released under NTU S-Lab License 1.0, which restricts redistribution and commercial use, and hosted deployments of the model carry the restriction forward — Replicate’s model page states its API cannot be used commercially.

Practitioner consensus is that GFPGAN holds up better on severely degraded old faces and CodeFormer preserves identity better on lightly damaged modern portraits. That is workshop experience rather than a measured result; the published literature treats the quality-versus-identity trade-off as unresolved for both.

Why this matters commercially: several paid tools in this category are thin interfaces over these three models. If you can run a GPU, you are paying for convenience, not capability.

11. Photomyne — the scanning step, priced opaquely

Photomyne’s value is bulk-scanning a physical shoebox with a phone camera, cropping multiple prints per shot, then colourising. Pricing is data not available in any usable form: third-party sources this run reported ten in-app purchases across eight price points from $1.99 to $59.99, a separate 10-year plan at $199.99, and standalone Colorize, SlideScan and FilmBox apps at $39.99 to $47.99 as multi-year one-off payments. The vendor’s own guidance is that the price depends on country, platform, trial eligibility and current store offers. A three-day free trial converts to a paid subscription automatically.

Best for: the digitisation step, if you accept that you cannot know the price in advance.

12. Google Photos — free, and not actually a restorer

Google Photos is included to correct a common assumption. Its Enhance control applies automatic tonal adjustments and will visibly improve a photograph that is simply too dark or washed out. It does not repair damage. Photo Unblur is free for all users with no monthly cap, while Magic Editor gives all Android and iOS users 10 saves a month; going beyond that requires a Pixel device or a Google One Premium plan at 2TB or above, so the 100GB and 200GB tiers do not help (Google).

Best for: tonal recovery on faded prints, for free, on photographs already in your library. Not for damage.

What restoration costs: AI against a human retoucher

The price gap is the largest in any category we cover, and it is worth stating before the accuracy section complicates it. The per-photo figures below come from restoration vendors’ own published comparisons and from studio price pages — that is, from parties with an interest in the answer — not from an independent survey. Treat them as indicative of shape rather than precise.

The structural point is not in dispute even if the figures are: AI pricing is flat regardless of how badly damaged the photograph is, because the model does the same work either way, while human pricing scales with damage.

Where the human is still worth it: severe damage, a photograph with only one surviving copy, and any case where the reconstruction has to be defensible. A retoucher can tell you what they inferred and why. A model cannot.

The accuracy problem: what AI restoration invents

This is the section the affiliate roundups leave out, and it is the most important thing on the page.

Restoration is generation. A tool that repairs a tear is not recovering hidden information — the information is gone. It is generating a plausible replacement. The better the model, the more plausible and therefore the harder to detect the invention becomes.

The independent evidence. The Second Challenge on Real-World Face Restoration at NTIRE 2026, run as a CVPR 2026 workshop challenge, is the only current independent benchmark in this area. It drew 96 registrants, 10 teams submitting valid models and 9 valid final scores, evaluated on a 450-image test set drawn from CelebChild-Test, LFW-Test, WIDER-Test, CelebA and WebPhoto-Test. The design detail that matters for buyers: entries were scored on a weighted image-quality assessment score and separately checked with the AdaFace identity model, because in this task visual quality and identity fidelity pull against each other. The winning entry, MiPlusCV, combined OSDFace-based coarse restoration with a one-step detail enhancement stage.

The specific failure modes are documented. The blind face restoration literature records that GFPGAN, DifFace and CodeFormer can generate artificial spectacles that were not in the source, and that VQFR, RestoreFormer++ and DAEFR can hallucinate facial features. Work on correcting this, such as DiffMAC, exists precisely because methods in this family tend to land on either identity distortion or incomplete restoration rather than solving both.

No consumer product on this page enters that benchmark. Remini, MyHeritage, VanceAI, Fotor and Hotpot publish no identity-fidelity figures of any kind, and there is no independent 2026 test scoring consumer restoration apps against each other on identity accuracy. Any “we tested 12 apps and ranked them” page you find for this keyword is scoring how good the output looks, which is the metric the research says is misleading. We are stating that rather than adding a thirteenth unverifiable test — which is also why the picks on this page rest on price, licence, capability scope and published limits, with practitioner consensus labelled as consensus wherever it is used.

The practical rule: keep the original scan, always, unedited, alongside the restored version. The restored file is an interpretation. The scan is the record.

Colourisation is a guess, and the research says so

Colourisation is not an unusually hard version of restoration. It is a different kind of problem.

It is mathematically ill-posed. Multiple colour images are equally valid outputs for the same greyscale input. There is no correct answer recoverable from the file; the model is picking a plausible one from what it saw in training.

Historical images are the hardest case. Research on colourising old photographs, including the HistoryNet model introduced in Focusing on Persons: Colorizing Old Images Learning from Modern Historical Movies, reports that existing methods do not perform well on historical person colourisation. HistoryNet addresses this with a classification sub-module covering era, nationality and garment type — the categories a model needs in order to know what a garment of that period would plausibly have been — and reports its gains on military uniforms — where historically correct colours are documented and can be checked — which tells you what the rest of the category is doing on garments nobody catalogued.

Era and culture are invisible to the model. Colours popular in a specific decade, traditional dress, ceremonial objects and culturally specific items are routinely rendered in generic tones. The model does not know your grandmother’s dress was green; it predicts a colour that fits images like this one.

The professional position. The critique is well established outside the tool industry — see Hyperallergic on the limits of AI colourisation of historical images and York University’s Hellenic Heritage Foundation Greek Canadian Archives on AI and historical photographs, published 13 March 2026, which argues for reading AI-produced historical images critically against other primary sources rather than at face value. The archival consensus in 2026 is that a colourised image should be labelled as a modern interpretation, not circulated as the document.

What to do about it: colourise for enjoyment, label it, keep the monochrome original, and where you do know a real colour, use a tool such as Palette.fm that lets you tell it.

Provenance: your restored photograph is now labelled as AI-edited

This is new since last year and almost nobody in this category has written about it.

Google’s marks. Images produced through Gemini carry two things: a visible sparkle mark and the invisible SynthID signal. Google made the visible mark optional in August 2026 via a Settings entry for media watermarking (TechCrunch, 14 August 2026). The invisible SynthID watermark and C2PA metadata are not affected by that toggle. Google describes SynthID as surviving screenshots, filters, compression, rescaling, recolouring and re-posting.

Google Photos’ labels. Google Photos adds Content Credentials to a file when an AI edit is saved, and where a file already carries valid credentials, subsequent edits append to them. The app surfaces this under “How this was made”, splitting provenance into media composition, AI edits, possible AI edits and non-AI edits (Google Photos Help). Non-AI edits such as a crop are recorded too. Note that this display is available on Android and iOS rather than on Google Photos on the web.

Why this cuts both ways. For a family archive, provenance is a benefit: in thirty years someone will be able to tell which file is the scan and which is the reconstruction, which is exactly what the archivists are asking for. For journalism, legal evidence, insurance claims, historical publication or anything where a photograph functions as proof, a restored file now carries a durable, machine-readable declaration that a generative model touched it — and that is the correct outcome, because one did.

The practical rule: if the photograph will ever need to function as evidence, do not run it through a generative restorer at all. Scan it, keep it, and have a human retoucher work on a copy. Our guide to AI watermarking explains how these signals work in more detail.

Feature and pricing comparison

Prices in USD. Where a figure is marked reported or disputed, it comes from third-party trackers rather than a vendor page we read directly, and sources disagree.

Best AI photo restoration tool for each need

For a torn, cracked or water-damaged print

Winner: Google Gemini

Generative regeneration beats patch-based repair on structural damage. The free tier runs Nano Banana 2; a Google AI plan at $19.99 a month unlocks Nano Banana Pro for the hardest cases. Alternative: Adobe Photoshop Generative Fill if you need to control the repair region by region.

For a face that has gone soft or grainy

Winner: Remini

Face reconstruction on degraded sources is its specialism and the easiest workflow on a phone. Alternative: CodeFormer with a high fidelity weight if you can run it locally and the non-commercial licence permits your use.

For colourising black-and-white family photographs

Winner: MyHeritage

Not because the colourisation is technically best, but because the result stays attached to the person, the tree and the source, and a photo-only subscription exists so you need not buy genealogy features you will not use. Alternative: Palette.fm if you want to steer the colours and you know some of them.

For restoring a hundred photographs at once

Winner: The open-weight pipeline on your own GPU, free and unlimited

Bringing Old Photos Back to Life is MIT-licensed and handles physical damage at any volume for nothing. Alternative: VanceAI if you want a hosted interface — but open its pricing page and confirm both the plan cost and the credits charged per restoration before committing, because published figures for both conflict.

For anything you will not upload

Winner: Topaz Photo, local processing

Its recovery and face recovery models run on your machine with no upload. Alternative: the open-weight stack, free and entirely offline.

For one photograph, right now, for nothing

Winner: The Google Gemini free tier

Nano Banana 2 on a free account is the strongest free restoration available, at a reported allowance of roughly 20 images a day, with the trade-off that output carries SynthID and C2PA provenance data. Alternative: Fotor, which needs an account and watermarks free exports, but is faster for a light fade.

For a photograph that has to remain evidentially faithful

Winner: None of these. Use a human retoucher on a copy, and keep the scan.

Every tool on this page invents pixels, and the Google tools attach a machine-readable AI-edit declaration to the file. This is the one use case where the correct answer is to buy a person, not a subscription.

For digitising a shoebox before restoring anything

Winner: Photomyne, price unclear

Multi-print capture from a phone is genuinely faster than a flatbed for volume. Alternative: a flatbed scanner at 600 DPI, which produces better source files and costs nothing per photo.

For a faded 1970s print with a colour cast

Winner: Google Photos Enhance, free

Tonal recovery is the one restoration job genuinely solved for free. Alternative: Lightroom’s auto tone if you already pay for it.

Before you restore: the scan decides the result

Every tool on this page is limited by what you feed it, and most disappointing restorations are scanning failures rather than model failures.

Scan at 600 DPI minimum for a standard print. More is better for small originals; a wallet-sized photograph at 300 DPI gives the model almost nothing to work with.

Scan flat, not photographed at an angle. Perspective distortion and uneven lighting are treated by the model as facts about the scene and preserved into the output.

Do not apply the scanner’s own “restore colour” or descreening options before the AI pass. They discard information the model could have used and add artefacts the model will faithfully reconstruct.

Keep the raw scan as a TIFF or maximum-quality JPEG, and run restoration on a copy. This is the only step on this page that is free, takes no skill and cannot be undone if you skip it.

How we rank

This is a restoration ranking, so we weight the four jobs the category actually covers — damage repair, face reconstruction, colourisation and tonal recovery — plus published price transparency, free-tier usefulness, whether the tool runs offline, licence terms for anything you might use commercially, and what happens to the file’s provenance metadata. We do not run our own visual quality test in this category, because the published research is explicit that perceptual quality and identity fidelity diverge, and a subjective visual comparison would score the wrong thing; where we rely on practitioner consensus instead, we say so in the line where we use it. On pricing, this category is the least transparent we cover: several vendors run region-dependent pricing, at least three publish no coherent price list, and in several cases vendor pages and third-party trackers gave us materially different numbers. We publish the conflict rather than pick a side, and mark a figure “data not available” rather than publish one we cannot stand behind. This page is re-scored monthly, and the date at the top is the last full pass.

Frequently asked questions

What is the best AI photo restoration tool in 2026?

For a physically damaged print, Google Gemini is the best option in September 2026. Free accounts run Nano Banana 2, which has been Google’s default image model since 26 February 2026, and Google AI Pro and Ultra subscribers keep the stronger Nano Banana Pro, regenerated from the three-dot menu. Both regenerate the damaged region as part of a coherent scene rather than cloning nearby pixels into the gap, which is why they handle cracks and tears crossing a face better than dedicated restoration apps built on older models. For a face that has gone soft rather than a print that has been torn, Remini is the specialist. For colourisation inside a family archive, MyHeritage fits best because the result stays attached to the person and the source.

Is Nano Banana Pro free?

No. Google says Pro and Ultra subscribers keep access to Nano Banana Pro for specialised tasks, by regenerating an image from the three-dot menu. Free Gemini accounts run Nano Banana 2, the Gemini 3.1 Flash Image model that became Google’s default across the Gemini app, AI Mode and Lens on 26 February 2026, at a reported allowance of roughly 20 images a day. Nano Banana 2 is the model most people restoring a photograph in Gemini are actually using, it is fast, and it is good enough that the free tier is still the strongest free restoration option in this guide. Google AI Pro is $19.99 a month.

Can AI really restore an old photo, or is it making things up?

It is making things up, and that is not a criticism — it is how the technology works. The information destroyed by a tear or lost to blur is gone from the file, so a restoration tool generates a plausible replacement rather than recovering hidden detail. The research on blind face restoration documents specific inventions, including spectacles that were not in the source and altered facial features, and the NTIRE 2026 face restoration challenge ran a separate identity checker alongside image-quality scoring because the best-looking output is often not the same person. Always keep the original scan unedited alongside the restored version.

What is the best free AI photo restoration tool?

The Google Gemini free tier, running Nano Banana 2, produces the best free results on serious damage, at a reported allowance of roughly 20 images a day — with the trade-off that output carries an invisible SynthID watermark and C2PA provenance data. Fotor is quicker for a light fade but is not as free as it is usually described: independent reviews report that it requires an account, grants a finite number of starter credits and watermarks free exports. If you can run software locally, Microsoft’s Bringing Old Photos Back to Life is MIT-licensed, free, unlimited and offline.

How much does it cost to have an old photo restored?

An AI tool costs between nothing and about $0.50 per photograph, a freelance retoucher typically charges $15 to $30, and a professional studio charges from around $30 for light damage up to $300 or more for a severely damaged print. These figures come from restoration vendors’ published comparisons and studio price pages rather than an independent survey, so treat them as indicative. The structural difference holds regardless: AI pricing stays flat no matter how damaged the photograph is, because the model does the same work either way, while human pricing scales with the difficulty of the job.

Is AI colourisation of old photos accurate?

No, and it cannot be. Colourisation is mathematically ill-posed: many different colour images are equally valid outputs for the same black-and-white input, so the model picks a plausible option rather than the correct one. Research on historical colourisation finds that general methods fail particularly on clothing, because they lack the semantic understanding to know what a garment of that era would have been, and era-specific, traditional and ceremonial items are routinely rendered in generic tones. Archival practice in 2026 is to label a colourised image clearly as a modern interpretation and keep the monochrome original as the record.

Does Photoshop have a photo restoration tool?

Photoshop has a Photo Restoration Neural Filter, but it has been unreliable. Adobe’s own community bug thread, opened in December 2025, records it returning “We’ve temporarily disabled this filter because of an error” on Photoshop 26.11.2 and 27.1.0; an Adobe community manager confirmed on 15 January 2026 that the fault still reproduced in versions 27.2 and 27.3 beta, and the thread was never marked resolved. Its current status is data not available, so test it on your install before planning around it. Photoshop’s generative tools — Generative Fill on a selected tear, the Remove tool on tape marks — are unaffected and are what most professionals actually use for restoration, because they let you control exactly which pixels the model may change.

Will a restored photo be marked as AI-generated?

If you restore it with Google’s tools, yes. Images produced through Gemini carry an invisible SynthID watermark and C2PA metadata; Google made only the visible sparkle mark optional, in August 2026. Google Photos adds Content Credentials to a file when an AI edit is saved and displays them on Android and iOS under a “How this was made” section covering media composition, AI edits, possible AI edits and non-AI edits. For a family archive this is useful, because it tells a future reader which file is the scan and which is the reconstruction. For a photograph that has to work as evidence, it is a reason not to use a generative restorer at all.

What is the difference between photo restoration and photo upscaling?

Upscaling increases resolution: it takes a small or soft file and produces a larger one with more pixels. Restoration repairs damage and recovers appearance: cracks, tears, stains, faded colour, blurred faces. An upscaler applied to a cracked photograph will frequently produce a larger, sharper crack, because it has no concept of damage. Some vendors sell both, which is the main source of confusion in this category. See our best AI image upscalers guide if resolution rather than damage is the problem.

Can I use a free AI restoration model commercially?

Check each licence individually, because they differ and are frequently misreported. Microsoft’s Bringing Old Photos Back to Life is MIT-licensed, covering both code and pretrained model, so it is unrestricted. GFPGAN states that it is released under Apache License Version 2.0, but it builds on third-party projects and pretrained components carrying their own terms, so check what your pipeline ships. CodeFormer is released under NTU S-Lab License 1.0, which restricts redistribution and commercial use, and Replicate’s hosted version states it cannot be used commercially. “Open source” in this category routinely means the code is open and some component is not commercially licensed, which is the trap that catches businesses building restoration into a product.

What resolution should I scan an old photo at before restoring it?

Scan at a minimum of 600 DPI for a standard print, and higher for small originals such as wallet-sized photographs, because the model can only work with the detail present in the file. Scan flat rather than photographing at an angle, since perspective distortion and uneven lighting are treated as facts about the scene and preserved into the output. Do not apply the scanner’s own colour-restore or descreen options first: they discard information the model could have used. Keep the raw scan as a TIFF or maximum-quality JPEG and restore a copy.

Which AI tool is best for restoring a blurry face?

Remini is the usual consumer recommendation for face reconstruction on degraded sources and the easiest to use on a phone, though it caps output at 2x and states a 4096 x 4096 pixel ceiling. If you can run software locally, practitioners generally reach for CodeFormer with a high fidelity weight on lightly damaged modern portraits and GFPGAN on severely degraded old photographs, though the published literature treats that trade-off as unresolved. Be aware that all three reconstruct rather than recover: the output is a plausible sharp face, not the face that was in the original file, and the difference matters if you are trying to identify someone.

Do I need a subscription to restore old photos?

No. Three routes avoid one entirely: the Gemini free tier for damage repair, Hotpot.ai at roughly $0.10 per image with nothing recurring, and the open-weight models on your own hardware for free and unlimited work. VanceAI is reported to sell one-off credit packs, though its pricing and its credits-per-restoration figures both conflict across sources, so check before buying. The tools that require a subscription — MyHeritage, Adobe Photoshop, Topaz — are worth it for the surrounding product rather than for the restoration itself.

Related guides

For related comparisons, see the best AI image upscalers if your problem is resolution rather than damage, the best AI photo editors for working on undamaged photographs, the best AI image generators for creating images from prompts, our guide to AI watermarking for how SynthID and C2PA provenance work, and the best AI apps for the wider consumer landscape.

This ranking is re-scored monthly. Pricing in this category is the least transparent we cover: several vendors run region-dependent and device-dependent pricing, at least three publish no coherent price list, and vendor pages and third-party trackers frequently disagree — so every figure marked reported or disputed should be confirmed on the vendor’s own site before purchase. Prices in USD and current as of 17 September 2026.