What are the best AI tools for photo restoration before and after, and how do I make accurate before/after comparisons?

The Honest State of AI Photo Restoration in Late 2026

AI photo restoration before-and-after comparisons have become the default marketing format for nearly every tool on the market, from mobile apps like the ones Perfect Corp. tested in early 2026 to desktop suites such as ON1 Photo Raw 2026.4 and Aiarty Image Enhancer V3.11. The category is crowded, the results vary wildly, and a meaningful number of restored images are not actually restored — they are re-generated. How-To Geek ran a piece in 2025 arguing that AI photo restoration is erasing your family's real history, and PetaPixel published a separate review of ON1 Restore AI under the headline "ON1 Restore AI Turns Old Family Photos into Grotesque Nightmare Fuel." Both critiques are accurate within their scope: when a face is below a certain pixel threshold and no reference data exists, the model invents plausible facial geometry that was never captured on film.

Also worth reading: What Is the Best Photo Restoration Software for Old Photos in 2026? · Can Free AI Photo Restoration Really Bring Faded Memories Back to Life? · Can AI Photo Restoration Be Ethical? Navigating the Colorization Dilemma?

The technology works best in narrow bands. Light scratch removal, gentle denoising, and tonal balancing on already-readable scans usually succeed 80–95% of the time in informal testing shared across creator forums. Heavy reconstruction — adding missing eyes, reconstructing closed eyelids, or recovering faces from scans under 300 pixels wide — fails or fabricates roughly 40–60% of the time, depending on the model. Anyone evaluating before/after examples should treat marketing images with the same skepticism applied to stock photography: the gallery is curated, the failures are not shown, and the "before" is often a synthetically degraded version of an already-clean image.

How AI Restoration Actually Works Behind the Curtain

Most consumer tools in 2026 use a diffusion-based pipeline trained on paired datasets of clean and degraded images. The model learns the statistical relationship between a faded 1970s portrait and a properly exposed one, then applies that relationship to your scan. Generative fill features — the kind that can replace a torn corner — use larger generative models similar in architecture to Stable Diffusion and DALL-E, which means they are sampling from a learned distribution of "what faces generally look like," not retrieving a memory of your grandmother.

This is the central technical reason a before-and-after pair can look dramatic and still be wrong. The model has no way to know whether a missing tooth was ever present, whether a childhood scar existed, or whether the eyes were originally brown or hazel. It produces the most statistically likely outcome. For group photos, that guess is often close. For a single tight portrait, the guess can rewrite the subject's identity. Understanding this distinction is the single most useful thing a non-technical user can learn before spending money on a subscription.

Colorization sits on top of the same architecture. Earlier convolutional approaches, including the 2016 DeOldify-style models, used reference databases of historical photographs to assign plausible colors. Modern systems embed colorization inside the restoration network, which means a black-and-white portrait sent through a single-pass restoration may receive tonal correction, scratch removal, and a colorization guess simultaneously. That convenience is also a risk: if the source is meant to remain black-and-white for archival reasons, the AI does not always respect that intent.

Building a Trustworthy Before-and-After Comparison

Creating a comparison that is honest and useful requires more discipline than the marketing sites suggest. The first rule is to never let the tool crop, reframe, or rotate between the two images. Even a 2% scale change breaks pixel-level comparison and makes a mediocre restoration look transformative. The second rule is to capture the "before" and "after" from the same source file at the same resolution, then export both at the same DPI and color space. A surprising number of viral before/after pairs online are mismatched in this way and have fooled hundreds of thousands of viewers.

For video comparisons, the creator economy has standardized on side-by-side sliders and short-form reveal animations. The Show HN post referenced in our research described a free tool for producing exactly this kind of Before-After comparison video, and the workflow it outlines — import two stills, align them, add a draggable wipe, export as MP4 — is what most professional reviewers now use. Keep the reveal between 2 and 4 seconds for short-form platforms and 6 to 10 seconds for YouTube; longer reveals lose viewers before the payoff.

The third rule is metadata hygiene. Embed the original scan resolution, the restoration tool and version number, the date processed, and any manual edits in the image description or a visible caption. This step costs almost nothing and turns a casual social post into a reference artifact that other restorers can cite. It also inoculates the comparison against the most frequent accusation of fabrication.

Comparison of Major AI Restoration Tools in 2026

The market now has roughly four operating tiers. Mobile-first apps target casual users with one-tap presets, desktop suites target enthusiasts who want manual control, and a small set of research-grade tools target archivists. The table below summarizes the most-cited options based on the Perfect Corp., Fstoppers, PetaPixel, and SLR Lounge coverage from early 2026.

The Perfect Corp. comparison of 12 apps placed Remini and MyHeritage in the top tier for face recovery, with PhotoMyne scoring highest for integrated scan-plus-restore workflows. Fstoppers found ON1's Restore AI effective on tonal damage but prone to inventing detail in eyes and mouths. SLR Lounge noted that Aiarty's V3.11 update meaningfully improved low-light face restoration and that lifetime licensing deals routinely drop the price near $40. None of the three publications declared a single winner, and the honest reading is that the choice depends on how much control you want versus how much time you want to spend learning a UI.

Common Mistakes That Ruin Restoration Results

The most frequent mistake is scanning at too low a resolution. A 600 DPI scan of a 4×6 print produces a roughly 2400×3600 pixel image, which gives any restoration model enough data to interpolate intelligently. A 300 DPI scan of the same print produces 1200×1800 pixels and forces the AI to invent half the face. If you only have access to a phone scanner, the better mobile apps now use multi-frame stacking that approximates a higher resolution, but this still loses to a flatbed scanner on archival work.

The second mistake is over-sharpening. Several reviewers in 2026, including the Fstoppers piece on ON1, flagged that aggressive sharpening produces a waxy, plastic appearance on skin, especially in older subjects. The third mistake is using a colorization pass on a photo that should remain monochrome for historical accuracy. The fourth is trusting a single-pass output: running a restoration twice with different models and comparing the result surfaces inconsistencies that should be flagged rather than accepted. The fifth is failing to back up the original scan before processing — if the AI corrupts the file or overwrites metadata, the source is gone. A 2026 anecdote covered in the research corpus involved a user losing a folder of irreplaceable family photos when a desktop AI tool crashed during a batch run; the originals were only recovered because iCloud sync had run a few minutes earlier.

When to Restore, When to Leave a Photo Alone

Not every damaged photograph benefits from AI restoration. Photos where the subject's identity is critical — a one-of-a-kind ancestor portrait, a wedding photo of someone now deceased — should be handled more cautiously than a snapshot of a building or a landscape. The risk-reward calculation is straightforward: if you cannot accept any invented detail, run the restoration at a low strength setting, or do not run it at all. If the photo is one of many similar shots and the goal is to view a clean version, full-strength processing is fine.

Photos with high historical value should be scanned, stored in two locations, and processed in copies. Cloud-to-cloud sync remains the most reliable backup for most users, and the cloud storage market now includes mature options from Apple, Google, and dedicated photo services. Reverse image search tools such as TinEye and Lenso.ai can also help establish provenance before restoration, which matters when a photo may have documentary significance. Getty Images' 2026 partnership with Nvidia around generative AI explicitly excludes training on archival editorial collections, a boundary that small consumer tools do not always respect.

Cost, Pricing, and What to Expect to Spend

Pricing varies by roughly two orders of magnitude. Free tiers exist on most mobile apps and on web-based services, typically with watermarks, resolution caps near 2000 pixels on the long edge, or a queue delay. Paid mobile subscriptions run $5 to $15 per month, with annual plans 30–50% cheaper. Desktop suites like ON1 Photo Raw 2026.4 charge a one-time fee near $100 for the base product, with the Restore AI module sometimes bundled and sometimes sold separately. Aiarty's V3.11 lifetime license was being offered at 49% off during 2026 with additional coupon stacking reported by SLR Lounge, putting the effective price near $40.

For a one-time restoration of a small family album of roughly 50 to 200 photos, a lifetime desktop license or a single web-tool purchase is the most economical path. For ongoing work — genealogists, local historical societies, professional photo retouchers — a subscription with batch processing is almost always cheaper per image. The break-even point sits near 80 to 120 images per year depending on the tool. Anyone processing thousands of images should evaluate API access, which several providers now offer at a per-image cost between $0.05 and $0.30.

What to Watch Through the Rest of 2026 and Into 2027

The category is moving quickly in three observable directions. First, face-specific restoration models are improving on low-resolution inputs, with several vendors claiming measurable gains on sub-300-pixel faces in 2026 release notes. Second, provenance and authenticity tooling is being added: invisible watermarking, C2PA metadata, and detection of AI-modified regions. These features matter because they let a restorer declare what was changed and what was left alone. Third, colorization is converging with restoration into single-pass pipelines, which is convenient but removes the user's ability to keep an image monochrome for archival reasons. Tools that offer an explicit "do not colorize" toggle will become more common through 2027 as users push back on default behavior.

The honest summary for anyone considering a new workflow: AI photo restoration is genuinely useful for 70–80% of typical home photograph damage, occasionally disastrous for the remaining 20–30%, and always worth approaching with a backup, a side-by-side comparison, and a clear sense of which details are negotiable and which are sacred.