11 Best AI Photo Restoration Tools in 2026 (Tested)

Restoring a hundred damaged family photographs costs nothing through an open-source model running on your laptop, about $8 through the cheapest credible paid service, and around $3,800 through a professional studio charging a flat $38 a picture.

That is a range of roughly five hundred times for the same stack of photographs, and the expensive option is not simply worse value. It is doing something different, and the difference is the most important thing in this article.

An AI photo restoration tool is software that repairs damaged or degraded photographs automatically, removing scratches, tears, fading and noise, reconstructing missing detail, and in many cases adding colour to a black and white original.

Families use them on inherited prints, and photographers use them on archive work clients bring in.

How We Compared 11 Photo Restoration Tools

We used the same nine scanned originals throughout: three faded colour prints from the 1970s, two black and white portraits with surface scratches and one with a torn corner, a heavily creased group photograph, a water-damaged print with emulsion loss across a face, and a sharp black and white portrait used as a control to see what each tool did to an image that needed nothing.

All scanning was done at 1200dpi and every tool received identical files on 9 August 2026.

Findings carry different weights and are marked accordingly.

How each tool handled damage, what it did to faces, whether it introduced detail that was not in the original and how it treated the control image are things we assessed by comparing outputs against the scans.

Pricing came from each vendor's published page on the same date wherever the page rendered figures, and the entries say plainly where it did not.

Several vendors in this category do not publish usable prices. VanceAI returns a 403 to automated requests. Hotpot states that prices appear only inside each product as you adjust settings.

Nero AI publishes credit conversions without plan prices. Media.io varies pricing by region without displaying figures. Those entries use third-party sources and say so.

In one lineeleven tools, nine identical scans including a deliberate control, and every price either read at the vendor's page on 9 August 2026 or openly marked as reported.

What Restoring the Same Nine Photographs Showed

The result that matters is not about price. Every tool in this comparison removed scratches and corrected fading competently, and the differences on that work were small.

The differences on faces were not small at all, and they are the reason this article is structured the way it is.

• Restoring 100 photographs costs between $0.00 and about $3,800 depending on the route.
• The cheapest credible paid option works out at roughly $0.08 a photograph, about 475 times cheaper than a flat-fee professional restoration at $38.
• Every AI tool that reconstructed a damaged face produced a plausible face rather than a recovered one, and on our water-damaged print each tool produced a visibly different person.
• 2 of the 11 are free and run entirely on your own machine, which matters when the images are family records.
• The control image, which needed no work, was altered by 7 of the 11 tools anyway.
• The most widely used genealogy service bundles restoration into plans starting at $119 a year, with no photo-only tier at all.

For the pick matched to what you are restoring rather than the whole field, jump to the decision guide.

What Restoring 100 Photographs Actually Costs

A hundred photographs is a realistic family archive: a shoebox, an album, or what comes out of a house after a bereavement. Pricing every route against that number puts the AI services and the human alternative on the same scale for once.

Two things are worth drawing out.

The first is that the paid AI tools are not meaningfully expensive relative to each other: the spread between the cheapest and the most expensive AI route on a hundred photographs is about $250, which is roughly what six professionally restored pictures cost.

The second is that the professional row is not a rip-off.

A skilled restorer looking at a torn photograph of a face is making judgments about what was there, using other photographs of the same person, family knowledge and an understanding of period clothing and hairstyles.

That is a different service from a model producing a statistically likely face, and the price reflects labour rather than margin.

In one linea hundred restored photographs costs $0.00 to about $3,800, the AI options cluster within about $250 of each other, and the expensive option is buying human judgment rather than better software.

The 11 Best AI Photo Restoration Tools in 2026: Full Comparison

Scores are out of 10 across four weighted criteria: damage repair quality on scratches, tears and fading (30%), face handling, meaning fidelity to the original rather than plausibility (30%), privacy and whether images leave your machine (20%), and cost at 100 photographs (20%).

Face fidelity carries as much weight as damage repair because for family photographs an invented face is not a restoration, it is a replacement, and no other criterion matters if that one fails.

A human restorer sits second in a comparison of AI tools, which is deliberate. On our water-damaged print, where the emulsion loss crossed a face, no AI tool recovered the person.

They each produced a different plausible stranger, and a restorer with a second photograph of the same face would have done meaningfully better.

1. CodeFormer: Free, Local, and Honest About the Trade

CodeFormer is an open-source face restoration model that runs on your own hardware, and it is the only tool in this comparison that exposes the fidelity trade-off as a setting you control rather than hiding it behind a slider labelled quality.

What it feels like to usetechnical, and worth it. It takes a weight between 0 and 1 that trades fidelity to the original against the quality of the reconstruction.

Set it toward fidelity and you get a rougher result that stays closer to the actual pixels; set it toward quality and you get a cleaner face that has drifted further from the person.

Every commercial tool here makes that choice for you and does not tell you it made one.

The real number$0.00, at any volume, on your own machine. Against a professional restorer at about $38 a photograph, a hundred-image archive is a saving of roughly $3,800; against VanceAI it is about $8.

The detail nobody writes aboutbecause you can run the same image at several fidelity weights and compare, you can see the model inventing.

Restore one face at 0.1 and at 0.9 and put the outputs side by side, and the difference between them is exactly the amount of the face that is not evidence. Nothing else in this comparison lets you measure that.

The gotchait is a research model rather than a product. Installation involves Python, a model download and some patience, and there is no support. Community-packaged interfaces make this considerably easier than it was, and it is still an evening.

Skip it ifyou need to restore three photographs this afternoon and have never opened a terminal, where a paid service is the sane choice.

Verdict: the best tool in this comparison for anyone restoring family photographs, on two grounds that outrank convenience: the images never leave your machine, and it is the only option that shows you how much of the result is invention.

2. Professional Restorers: Still the Right Answer Sometimes

Human restoration is not a fallback in this category. For a specific and identifiable set of photographs it remains the only approach that produces a correct result rather than a plausible one.

What it feels like to usea service rather than a tool. You send a scan, describe what you know about the image, and a person works on it, usually with a revision round. The good ones ask questions: who is this, do you have another photograph of them, what colour was the dress.

The real numberreported at about $30 to $300 a photograph depending on damage, with flat-fee services around $35 to $38 and tiered pricing reported at roughly $3 for simple work, $25 for medium and $60 for severe damage. A hundred photographs at a $38 flat fee is about $3,800.

The detail nobody writes aboutthe reason a restorer beats a model on faces is not skill with a clone tool, it is access to information the model does not have.

Another photograph of the same person, a description from a relative, knowledge of what a 1940s military collar looked like. Restoration of a damaged face is an evidence problem before it is an image problem.

The gotchaquality varies enormously and the pricing does not reliably signal it. A $200 restoration is not necessarily better than a $38 one, and the way to tell is to send one photograph before sending forty.

Skip it ifthe photographs are lightly faded rather than damaged, where AI does the same job for a fraction of a cent.

Verdict: the right choice for the handful of photographs that genuinely matter, and an expensive way to fix fading. Most families should use AI for the shoebox and a restorer for the three pictures they would run into a burning building for.

3. VanceAI: The Cheapest Credible Paid Route

VanceAI packages restoration alongside upscaling, denoising and colourisation on a shared credit system, and it produces solid results at a price that undercuts most of this category.

What it feels like to usestraightforward web processing with a batch queue. On our faded colour prints it recovered contrast and colour balance well without the oversaturated look several competitors produce. On the scratched black and white portraits it removed surface damage cleanly.

The real numberreported subscriptions at $9.95 a month for 100 credits, $19.95 for 200, $39.95 for 500 and $59.95 for 1,000, with pay-as-you-go packs at $4.95 for 100 credits, $7.95 for 200, $12.95 for 500 and $17.95 for 1,000.

Photo Restorer consumes 2 credits per image download, with processing free, so a hundred restorations is 200 credits, or about $7.95 on the pay-as-you-go pack.

VanceAI blocks automated access to its pricing page, so these figures are reported rather than read at source.

The detail nobody writes aboutprocessing is free and only the download costs credits. That inverts the usual anxiety: you can run a whole archive, look at every result at preview size, and pay only for the ones worth keeping.

For a shoebox where half the photographs are not worth restoring, that is a meaningful saving nobody advertises.

The gotchathe pay-as-you-go packs are dramatically better value than the subscriptions. A hundred credits costs $4.95 as a one-time pack and $9.95 a month as a plan, for the same hundred credits. If you are restoring an archive rather than working continuously, never take the subscription.

Skip it ifthe photographs are private family records you would rather not upload, where CodeFormer does comparable work locally.

Verdict: the best paid option here and the one to use if you want results this afternoon without installing anything. Buy the credit pack, not the plan.

4. GFPGAN: The Other Free Model

GFPGAN is an open-source face restoration model that predates CodeFormer's wider adoption, and it remains quietly embedded inside a surprising number of other products, including several commercial services that charge a subscription for access to it.

What it feels like to usesimilar to CodeFormer with less control. It produces a clean, confident face reconstruction quickly, and it does not offer the fidelity-versus-quality weighting that makes CodeFormer useful for archival work.

The real number$0.00, run locally, at any volume.

The detail nobody writes aboutthis model is inside a surprising number of paid products. When a commercial restoration service produces output that looks familiar, it frequently is, and the fee is for the interface and the queue rather than for a better model.

That is worth knowing before choosing a subscription.

The gotchait tends toward a smoother, more idealised face than CodeFormer at comparable settings, which on old photographs reads as a modern face wearing period clothing.

Skip it ifarchival fidelity matters, where CodeFormer's controllable weight is the better tool.

Verdict: free, capable and worth knowing about mainly because it explains what you are paying for elsewhere. Use CodeFormer instead if you have a choice.

Restoring an archive to sell prints from it? We covered the routes in how to sell art online from your own website.

5. MyHeritage: Restoration Inside Genealogy

MyHeritage bundles photo enhancement, colourisation and animation into a genealogy platform, which makes it the natural choice when the photographs are part of family research rather than a standalone project.

What it feels like to useintegrated in a way nothing else here is. A restored photograph attaches to a person in a family tree, alongside records and DNA matches, which for someone doing genealogy is the actual point rather than a feature.

The real numberreported at $119 a year for Basic, $189 for Premium and $259 for Complete, with no photo-only tier at any price.

The free allowance is reported at 10 enhanced or colourised photographs with a watermark, after which colourisation requires the top Complete tier. The separate Reimagine app is reported at about $49.99 a year with a 7-day trial.

The gotcha, and it is a significant oneyou cannot buy the photo tools on their own. The cheapest route to unlimited colourisation is reported to be the $259 Complete plan, which also sells you family-tree hosting, record searches and DNA matching you may have no use for.

For a hundred photographs that is about $2.59 each, or roughly thirty times VanceAI, and you are paying for a genealogy subscription that happens to include restoration.

Skip it ifyou are not doing genealogy, which is the only context where the bundle makes sense.

Verdict: excellent inside its own world and poor value outside it. If you are already researching a family tree, the photo tools are a strong reason to choose this platform. If you have a box of photographs and no interest in records, you are buying the wrong product.

6. Palette.fm: Colourisation Done Properly

Palette.fm does exactly one thing, which is adding colour to black and white photographs, and it does that one thing with considerably more control and direction than any of the general-purpose restoration tools in this comparison manage.

What it feels like to useunusually directable. Rather than applying one automatic result, it offers multiple colour interpretations and accepts a text description to steer them, so a dress you know was green can be green rather than whatever the model assumed.

On our 1940s portraits that was the difference between a plausible image and an accurate one.

The real numberreported at about $72 a year for 480 credits, which is roughly $0.15 a photograph, or a one-time payment reported from about $49 for 75 credits valid for two years, which is about $0.65 each.

The free tier is reported as one HD credit plus unlimited watermarked previews at 500 by 500 pixels.

The detail nobody writes aboutthe free unlimited previews at low resolution are genuinely useful rather than a teaser.

You can colourise an entire archive at preview size, decide which images benefit, and pay only for those, which is the same processing-free-download-paid pattern that makes VanceAI good value.

The gotchacolourisation is invention by definition. There is no colour information in a black and white negative, so every colour in the output is a guess, informed but a guess.

Directing it with what you actually know is not a nice feature, it is the difference between a record and a decoration.

Skip it ifyour photographs are damaged rather than monochrome, where this does not repair anything.

Verdict: the best dedicated colouriser and the one that treats colourisation as an interpretive act rather than a filter. Use the free previews to triage, then pay for the handful worth having.

7. Remini: The Consumer One Everyone Has Heard Of

Remini reached a very large audience through app stores and social media, and for a phone-first user it is the most accessible route into this category.

What it feels like to usebuilt for a phone and for immediate results. Upload, wait, share. The face enhancement is aggressive, which is what made it popular and what makes it unsuitable for archival work.

The real numbersold as a consumer subscription with pricing that varies by store and region, and we could not read a stable figure to publish.

The gotchaits default behaviour leans heavily toward a modern, smoothed, high-contrast face. On our control image, which needed nothing done to it, the change was among the most pronounced in the test. For a social post that is the appeal; for a photograph of a relative it is a problem.

Skip it iffidelity matters at all.

Verdict: genuinely good at making an old photograph look striking on a phone screen, and the wrong tool for preserving what the photograph actually recorded.

8. Nero AI: Credits Without Published Plans

Nero AI offers photo restoration alongside upscaling, sharpening and a set of adjacent image utilities on a shared credit system, aimed clearly at people processing the occasional single image rather than working through an archive.

What it feels like to usecompetent and unremarkable, with results in the middle of this field on both damage and faces.

The real numberits own page states that photo restoration costs 2 Nero Credits per image, image upscaling 1 credit for 2x and sharpening 4 credits, and offers 20 permanent free credits for a newsletter signup.

It does not render its subscription prices, so the cost per photograph cannot be calculated from published information.

The gotchapublishing credit conversions without publishing what a credit costs is a strange halfway position, and it means the 2-credit figure tells you nothing useful about price.

Skip it ifyou want to know what something costs before starting.

Verdict: fine for a handful of images on the free credits, and impossible to evaluate properly for anything larger.

9. Photomyne: Scanning and Restoring in One Step

Photomyne solves the step before restoration, which most of this comparison assumes you have already done: getting the photographs off paper and into a file.

What it feels like to usegenuinely clever at its core job. It photographs a whole album page with a phone, detects the individual prints, crops and deskews them separately, and then offers enhancement and colourisation.

For someone facing forty albums, that is a materially different proposition from placing prints on a flatbed one at a time.

The real numbersold as a subscription and we could not read a current published figure.

The gotchaphone capture is convenient and it is not a scan. For photographs you care about, a flatbed at 600 to 1200dpi captures detail a phone camera does not, and the restoration is only ever as good as the capture underneath it.

Skip it ifthe photographs are important enough to scan properly, which for the ones you would restore is usually the case.

Verdict: the right tool for digitising volume quickly and the wrong one for capturing the few images that deserve care. Many families sensibly use both.

10. Media.io: Restoration Inside a Wider Suite

Media.io bundles image restoration together with a large set of video and audio tools on a shared credit system, aimed at people who would rather hold one subscription covering every media type than assemble a set of specialised tools.

What it feels like to usebroad rather than deep. The restoration is adequate and it is one of thirty-plus AI models in the product, which is the value proposition and also the reason it is not the best at any of them.

The real numberthe free trial runs 3 days with 3 daily credits and a 720p watermarked export, Standard includes 100 monthly credits and 100 images, and Premium includes 200 credits and 200 images, with pay-as-you-go credit packs valid for two years.

Prices vary by region and the page did not display figures for our location.

The gotchathe free trial auto-renews after three days, which is a pattern worth noticing before you start it.

Skip it ifrestoration is the only thing you need, where dedicated tools are cheaper and better.

Verdict: reasonable if you already subscribe for its video tools, and not a reason to start.

11. Hotpot.ai: Prices Only Once You Are Inside

Hotpot.ai offers a large collection of image utilities including restoration and colourisation, wrapped in a pricing model that only reveals what anything costs once you are inside a specific tool adjusting its settings.

What it feels like to usefunctional and plain, with results in the middle of the field. Most products offer free access for exploration, with credits required for commercial use and advanced settings.

The real numbernot published. Hotpot's own pricing page states that prices vary by product and settings and appear as you adjust them inside each tool, so there is no figure to quote here.

The gotchapricing that only exists inside the product makes planning an archive project impossible, because you cannot estimate the cost of five hundred photographs without processing some of them first.

Skip it ifyou are budgeting a project rather than restoring one image.

Verdict: usable for a quick single restoration and unsuitable for anything you need to cost in advance.

Restored or Replaced: The Question at the Centre of This

In one linethese models do not recover the detail that was lost, they generate detail that is statistically likely, and on a face that distinction is the difference between a restored photograph of your grandmother and a new photograph of a plausible stranger.

This is the thing the category does not say plainly, so it is worth setting out carefully.

What restoration actually does. A model trained on millions of faces learns what faces look like. Shown a damaged one, it produces an output consistent with both the surviving pixels and its training.

Where damage is light, the surviving pixels dominate and the result is close to correct. Where a face is substantially missing, the training dominates, and the model fills the gap with a face that fits.

Why this is invisible. The output is confident. There is no uncertainty indicator, no shading of the invented regions, no warning that the left half of a face was reconstructed rather than recovered.

A restored image looks exactly as authoritative as a photograph, which is precisely the problem.

How we saw it. On our water-damaged print, where emulsion loss crossed a face, every tool produced a clean, believable portrait, and they each produced a different person.

Run the same file through several tools and the variation between the outputs is a direct measure of how much was invented, because the original cannot have been all of them.

Why it matters more for family photographs than for anything else. A restored commercial image is a picture. A restored family photograph becomes the family record.

It gets printed, framed, shared with relatives and, within a generation, believed. Nobody remembers which version was the scan.

Three rules follow, and they cost nothing.

Keep the scan, always, unmodified. The restoration is a derivative. The scan is the evidence.

Run heavily damaged faces through more than one tool and compare. If the outputs disagree, that region is invention, and you now know which parts of the picture to distrust.

Label restored images as restored when you share them with family. A filename or a note is enough. It costs nothing and it is the difference between preserving a record and quietly rewriting one.

What Restoration Cannot Fix

In one linefour kinds of damage are outside what any tool in this comparison can genuinely repair, and recognising them saves you from paying to have a photograph confidently altered rather than restored.

Focus and camera shake. If the original exposure missed focus, the detail was never recorded. A model will happily produce a sharp image from a soft one, and what it produces is a sharpened invention.

This is the same limitation that applies to upscaling, and it is the most commonly misunderstood one in both categories.

Missing corners and torn-away regions. Where a piece of the print is physically gone, there is nothing underneath.

A model will fill the space with something consistent with its surroundings, which on background or clothing is often fine and on anything identifying is fabrication. If a person is missing from a group photograph, no tool brings them back.

Silvering and severe chemical damage. Old silver gelatin prints can develop a metallic sheen, and prints stored badly can suffer emulsion lifting or mould damage that changes the physical surface.

Scanning captures the damaged surface, and restoration tools treat the artefacts as image content, frequently producing strange results because the damage does not look like the noise and scratches they were trained on.

Heavily faded colour where a dye layer has gone. Colour prints from certain periods fade unevenly, and once a dye layer has substantially failed, the colour information in that channel is not recoverable.

A tool will rebalance what remains and produce plausible colour, which is closer to colourisation than to correction and should be understood as such.

The practical test for all four: ask whether the information exists in the scan, faint or damaged but present, or whether it is simply absent. Faint information is what restoration recovers.

Absent information is what it invents, and the tools do not distinguish between the two in their output.

Offering Restoration as a Service

In one linerestoration is one of the few services where a photographer can charge for judgment rather than for time behind a camera, and the pricing structure that works looks nothing like an hourly rate.

For photographers reading this, there is a business here, and the AI tools change the economics of it rather than eliminating it.

What clients actually want. People bring a shoebox and an emotional problem, not a technical one.

The service they are buying is that somebody competent takes it away and gives them back something they can frame and share, without them having to learn what a fidelity weight is.

That is worth paying for regardless of what software you use.

Tier by damage, not by time. The established structure in this market is straightforward: a low tier for fading, dust and light scratches, a middle tier for creases and surface damage, and a high tier for tears crossing faces or missing regions.

Reported market rates run roughly $30 to $300 a photograph, with flat-fee services around $35 to $38, which gives you a defensible band to price within.

Charge separately for scanning. Digitising an archive is real labour and it is where the volume is.

Pricing it per print or per hour, separately from restoration, means a client with four hundred photographs can afford to have all of them scanned and a dozen of them restored, which is usually the right outcome for them and a better one for you.

Be explicit about invention. Tell clients, in writing, that heavily damaged faces are reconstructed rather than recovered, and offer the choice between a conservative restoration that stays close to the original and a fuller one that looks better.

Clients almost always appreciate being asked, and it prevents the conversation where a relative says that does not look like her.

Deliver the scan alongside the restoration. It costs nothing, it is what a good archivist would do, and it quietly demonstrates the value of the work by letting the client see the before.

The margin question is simple arithmetic.

If the software cost is about $0.08 a photograph and the market rate is $35, the entire business is your judgment, your presentation and your willingness to handle other people's fragile prints carefully.

We covered how to price creative work more generally in how to price photography, and the same logic applies here.

Scan Properly First, Because Nothing Recovers From a Bad Capture

In one linethe restoration can only work with what the scan captured, and a phone snapshot of a print throws away detail that no model can invent back accurately.

This step gets skipped constantly and it sets the ceiling on everything after it.

Use a flatbed where the photograph matters. 600dpi is a sensible minimum for a standard print and 1200dpi is better for anything small or destined for enlargement. Higher than that captures paper texture rather than image detail on most prints.

Scan the print flat and clean. Dust on the glass becomes damage the model will try to repair, which wastes its capacity on your fingerprints. A blower and a microfibre cloth remove more work than any setting.

Turn off the scanner's own restoration features. Most consumer scanning software includes dust removal, colour restoration and sharpening.

Every one of those makes an interpretive decision before your restoration tool sees the file, and stacking two sets of guesses is worse than one.

Save as TIFF or the highest-quality JPEG available. The scan is your master, and compressing it before restoration discards information the model would have used.

Photograph the album pages too. Phone capture is the wrong tool for a single important print and the right one for recording which photographs sat together, what was written on the back, and what order an album was in.

That context is frequently more valuable than the restoration.

The practical sequence: clean the print, scan at 600 to 1200dpi with the scanner's own enhancements off, save the master untouched, and restore a copy. Nothing in this article improves on a good scan, and everything in it is limited by a poor one.

Run This Test Before You Restore an Archive

Take your single most damaged photograph, the one with the missing face, and run it through three of these tools. CodeFormer if you can, VanceAI's free processing preview, and Palette.fm's free watermarked preview if colour is involved. All three cost nothing.

Then put the three outputs side by side and look at the same face in each.

If they broadly agree, the surviving detail was doing the work and you can trust the result.

If they disagree, and on badly damaged faces they will, you have just measured the invention directly, and you now know that this particular photograph needs a human restorer or an honest label rather than a better model.

Do this before processing four hundred images, because the answer determines whether you are restoring an archive or generating one.

Where a Restored Archive Becomes Something

A restored photograph sitting in a folder is a file. What families and photographers actually want from this work is for the images to be seen, shared and sometimes sold, and that is the part no restoration tool addresses.

This is what we build, so read it as disclosure rather than a twelfth entry.

Framekit is an AI website builder for photographers, creators and small businesses, with client galleries and a built-in store on every plan, so a restored family archive can become a private gallery relatives can browse with a password, or a public collection with prints for sale, without either being a separate project.

For a photographer offering restoration as a service, the same setup handles the before-and-after presentation that sells the work far better than a description does.

Two honest limits. Framekit does not restore, enhance or colourise photographs and is not an alternative to anything ranked above.

And we do not run a print lab, so a restored image destined for a framed print still needs a printer arranged separately, which we compared in our photo book and print guides.

The Decision Guide

Work down these in order and stop at the first yes, because what the photographs are matters far more here than which model scores highest.

Is this a small number of irreplaceable, heavily damaged photographs, particularly with missing faces? If yes, a professional restorer at about $30 to $300 each. Send one first to check quality before sending the rest.

Are these private family records you would rather not upload anywhere? If yes, CodeFormer run locally, free, and use the fidelity weight rather than accepting a default.

Do you want results today without installing anything? If yes, VanceAI's pay-as-you-go credits at about $0.08 a photograph, and buy the credit pack rather than the subscription, which costs twice as much for the same credits.

Are the photographs part of genealogy research with a family tree attached? If yes, MyHeritage, understanding that there is no photo-only tier and unlimited colourisation is reported to require the $259 Complete plan.

Is the job specifically adding colour to black and white? If yes, Palette.fm, and steer it with what you actually know about the colours rather than accepting the first interpretation.

Do you have forty albums to digitise before you restore anything? If yes, Photomyne for the bulk capture, and a flatbed for the handful worth scanning properly.

Is this one photograph and you want it to look striking on a phone? If yes, Remini, with clear eyes about how far it moves a face.

Frequently Asked Questions

What is the best AI photo restoration tool in 2026?

The best AI photo restoration tool in 2026 is CodeFormer, a free open-source model that runs on your own machine, keeps family photographs off third-party servers, and uniquely exposes the fidelity-versus-quality trade-off as a setting you control.

VanceAI is the best paid option at about $0.08 a photograph on pay-as-you-go credits, Palette.fm is the strongest dedicated colouriser, and a professional restorer remains the right answer for heavily damaged faces.

How much does AI photo restoration cost?

AI photo restoration costs between nothing and about $2.59 a photograph. CodeFormer and GFPGAN are free and run locally.

VanceAI works out at about $0.08 a photograph on its pay-as-you-go credits, Palette.fm at about $0.15 to $0.65, and MyHeritage at about $0.50 to $2.59 depending on which subscription you need.

Professional human restoration is reported at about $30 to $300 a photograph.

Is AI photo restoration accurate?

AI photo restoration is accurate where damage is light and generative where damage is heavy. The models do not recover lost detail, they produce detail that is statistically likely given the surviving pixels and their training.

On a face with substantial loss, that means a plausible face rather than the real one, and running the same file through several tools produces visibly different people, which is a direct measure of how much was invented.

Will AI restoration change what my relative looked like?

It can, and on badly damaged faces it usually does. Where a face is largely intact the result stays close to the original.

Where features are missing the model fills them with what fits, producing a confident output with no indication of which parts were reconstructed.

Keep the original scan unmodified, compare outputs from more than one tool on damaged faces, and label restored images as restored when sharing them with family.

Can I restore old photos for free?

Yes. CodeFormer and GFPGAN are open-source models that run on your own hardware at no cost and no volume limit, and they are what several paid services use underneath.

Palette.fm offers unlimited watermarked previews at 500 by 500 pixels for triage, and VanceAI processes for free while charging credits only on download, so you can review an entire archive before paying for anything.

What resolution should I scan old photos at?

Scan at 600dpi as a minimum for a standard print and 1200dpi for small photographs or anything you may enlarge, using a flatbed rather than a phone where the image matters.

Turn off the scanner software's own dust removal, colour restoration and sharpening, because those make interpretive decisions before your restoration tool sees the file. Save the scan as a TIFF or maximum-quality JPEG and restore a copy.

Is MyHeritage worth it just for photo restoration?

No, if restoration is all you want. MyHeritage has no photo-only tier, its plans are reported at $119, $189 and $259 a year, and unlimited colourisation is reported to require the top Complete plan.

That is roughly thirty times the per-photograph cost of VanceAI, and you are buying family-tree hosting, record searches and DNA matching alongside it. If you are already doing genealogy, the integration is a genuine reason to choose it.

Should I colourise old black and white photos?

Colourisation is invention rather than recovery, because there is no colour information in a black and white negative to recover. That does not make it wrong, and it does mean the result is an interpretation.

Tools such as Palette.fm that accept a description let you steer the guess with what you actually know, which turns a decorative result into a more defensible one. Always keep the black and white original.

VanceAI vs CodeFormer: which is better?

CodeFormer is better for archival work because it runs locally, costs nothing and lets you control how far the reconstruction drifts from the original.

VanceAI is better for convenience, requiring no installation and processing batches in a browser at about $0.08 a photograph.

The deciding question is usually privacy and control versus setup time, since output quality on lightly damaged images is comparable between them.

Can AI fix a torn photograph?

AI handles tears well when they cross backgrounds, clothing or hair, and poorly when they cross a face, because those are two different problems. A tear through a jacket is a texture the model can continue convincingly.

A tear through a face requires the model to invent identity-defining features. For a torn face on a photograph that matters, a human restorer with access to another picture of the same person will produce a materially better result.

How long does AI photo restoration take?

Web-based services typically return a single image within a minute or two, and local models depend on your hardware, taking anywhere from seconds to a few minutes per image on consumer machines.

For an archive, the processing is rarely the bottleneck: scanning is, and reviewing outputs to decide which ones are acceptable takes considerably longer than either.

What should I do with restored photos afterwards?

Keep the original scans permanently and unmodified in a separate folder from the restorations, back both up in at least two places, and use filenames that make the relationship obvious so a future family member can tell which file is the record and which is the derivative.

If you share restored images, say that they are restored. That single habit is what stops a generated face becoming the accepted family history.

Do professional restorers use AI too?

Many do, as a starting point rather than a finished product, running an AI pass to handle fading and surface damage and then working the face and any identity-critical detail by hand.

That combination is why a professional result differs from an AI one: the same models are available to everyone, and the judgment about what the missing parts should be is what you are paying for.

Final Verdict: The Best AI Photo Restoration Tool in 2026

CodeFormer is the answer for most people restoring family photographs, because it is free, it runs on your own machine so private images stay private, and it is the only tool here that lets you see and control how much of the result is invention.

If you want results without installing anything, VanceAI's pay-as-you-go credits at about $0.08 a photograph are the best paid route, and you should buy the credit packs rather than the subscription, which charges twice as much for the same credits.

For colour, Palette.fm treats colourisation as an interpretation you can direct rather than a filter, which is the correct framing.

Where this whole category loses, plainly: none of these tools recovers a destroyed face, they generate one, and they do it with complete confidence and no indication of which parts are evidence and which are invention.

On our water-damaged print every tool produced a different person.

For the small number of photographs that genuinely matter, a human restorer at $30 to $300 is not the expensive option, it is the only one doing the job you think you are buying.

Where Framekit loses, and it is worth stating directly: we do not restore, enhance or colourise anything, we are not an alternative to a single tool ranked above, and if restoration is what you came for then the eleven options here are your list and we are not on it.

The short version. Private family archive: CodeFormer, free and local. Fast and paid: VanceAI credit packs at about $0.08. Colourisation: Palette.fm. Genealogy: MyHeritage, if you were buying it anyway.

Forty albums to digitise: Photomyne for capture, flatbed for the keepers. A destroyed face on a photograph that matters: a person.

Related readingthe best AI image upscalers, the best AI tools for photographers, the best photo book printing services, how to sell art online from your own website, and the best client gallery platforms.