How to restore old photos with AI in 2026?

What "AI Restoration" Actually Means in 2026

Restoring an old photo with AI in 2026 is no longer a single operation. The term now covers at least five distinct technical tasks: scratch and tear removal (inpainting), denoising, face restoration, super-resolution upscaling, and colorization. Most consumer apps bundle several of these into one button, which is convenient but masks what is actually happening. The model is guessing missing pixels from patterns it learned during training, not recovering the original data, because the original data no longer exists. As PetaPixel argued in 2025, calling this process "restoration" is a broken concept when the source file is gone. The realistic framing is "AI-assisted reconstruction," and the output quality depends heavily on the input scan quality and which model is being used.

Also worth reading: How do I use an AI colorization reference image workflow to restore historical photos accurately? · Can AI really restore old family photos without ruining them, and what should you know before trying it? · What are the ethical guidelines for using AI to restore and colorize archival photos and footage in 2026?

A useful mental model: think of the AI as a very fast junior retoucher. It will get about 80% correct in seconds, and the remaining 20% is where your judgment, manual editing, or a human professional still matters. Anyone promising 100% automated perfection on a torn, faded, water-damaged print is overselling what the current technology can do.

How the Underlying Technology Works

Modern photo-restoration pipelines stack three or four neural networks. First, a super-resolution model (often built on the ESRGAN or Real-ESRGAN architecture used in tools like ON1 Photo Raw 2026.4) upscales the image, sometimes by 4x or more, to add plausible texture. Second, a face-restoration network, frequently CodeFormer or GFPGAN, identifies facial regions and reconstructs eyes, mouths, and skin at higher fidelity than the surrounding image. Third, a denoising diffusion inpainting model fills in scratches and creases by predicting what is statistically likely to be underneath. Colorization models such as DeOldify or MyHeritage's In Color add plausible hues based on a learned understanding of period-appropriate clothing, skin tones, and environments.

The catch is that each model was trained on different datasets, so they have different biases. CodeFormer, for instance, tends to smooth skin aggressively, which can make a 1962 portrait look like a 2024 beauty-shoot still. Real-ESRGAN can hallucinate fabric patterns that never existed in the original garment. Knowing which tool was used to produce a result helps you decide whether to trust it. The PicWish upgrade released in late 2024, for example, advertises an Ultra HD output model specifically because previous versions capped at resolutions that exposed artifacts on large prints.

Practical Step-by-Step Workflow

A sensible workflow in September 2026 starts long before you open any app. Scan the physical photo at 600 DPI or higher using a flatbed scanner rather than a phone camera, unless the photo is framed or too large to fit. Use a color target if you have one, because faded prints often have a strong color cast that will throw off the AI. Save a master TIFF as your untouched archive file. Never let the AI touch your only copy.

Once the scan is ready, run it through an AI restoration tool. For most users, MyHeritage in Color, Remini, or PicWish will produce acceptable results in under a minute. For more control, ON1 Photo Raw 2026.4 lets you dial the strength of each restoration pass and combine AI with manual masking. After the AI pass, open the result in Photoshop, Affinity Photo, or GIMP and spend another five to ten minutes cleaning up artifacts, sharpening eyes, and correcting any skin tone that looks uncanny. Finally, export at the largest size the source image can support, typically 4x to 8x the original pixel dimensions.

Comparison of Major AI Restoration Tools

The market in 2026 is crowded, and pricing models vary widely. Below is a snapshot of how the leading options compare based on publicly available information and recent reviews.

Perfect Corp's 2024 test of 12 restoration apps found a quality spread of roughly 30% between the best and worst performers on identical inputs, which is a meaningful gap when the goal is a large print. ON1's own marketing acknowledges that Restore AI can produce "grotesque" results on low-resolution inputs, a rare honest disclaimer in this space.

Colorization Is a Separate Discipline

Colorization deserves its own section because it produces the most visually dramatic results and the most disappointment. AI colorization does not retrieve lost color information; it invents plausible color based on learned patterns. A 1943 wedding photo will receive green grass, blue sky, and skin tones that match the average training image of that era, but those specific shades may be wrong. Historical uniforms, military insignia, and ethnic garments are frequent casualties.

For high-stakes historical work, colorization should be treated as a first draft. Researchers at the Monroe County Library system, who presented on AI photo restoration in 2024, recommend using AI colorization as a starting reference and then manually adjusting known elements like uniform colors, team jerseys, or flag hues. For casual family photos this is usually fine. For museum or archival work, it is not.

Common Mistakes That Ruin Results

The single biggest mistake is starting with a low-quality scan. A phone photo of a photo, shot at an angle with glare, gives the AI nothing to work with and produces mush. The second biggest mistake is over-processing. Running an image through three AI tools in sequence compounds artifacts and can erase legitimate detail. A 2024 PetaPixel test of ON1 Restore AI showed faces being warped into uncanny shapes when the model's strength slider was set above 70%.

Other frequent errors include: trusting the first output without comparing it to the original, failing to back up the source file, using JPEG output throughout the chain (each save loses data), and assuming AI restoration is free. The "free" tier of most apps is limited to a handful of images per day or watermarks the output, and unlimited use typically costs between $5 and $20 per month. Saving the restored image at low resolution for social media is fine; saving it that way as your only archive copy is not.

When Human Restoration Still Wins

For torn photographs, photos with missing sections, or images where a face has been physically destroyed, a skilled human retoucher using Photoshop still outperforms AI on most inputs. The cost is roughly $25 to $75 per hour, and a complex restoration might take two to four hours. AI gets you 80% of the way there for $0 to $10, and a human finishes the last 20% that the AI cannot. For sentimental family photos where the face of a deceased relative is partially gone, hybrid workflows are usually worth the extra cost.

The opposite is also true: AI has decisively won on tasks like bulk colorization of hundreds of slides, gentle scratch removal on otherwise intact prints, and upscaling small snapshots for printing at 8x10 or larger. A human retoucher would charge $10-$30 per slide for colorization; MyHeritage will colorize hundreds for the same monthly subscription. The economics are not close for volume work.

Cost and Value Considerations

Pricing varies sharply. Free tiers exist across nearly every major app, but they are either heavily watermarked, capped at 2-5 images per day, or limited to low-resolution output. Mid-range subscriptions run $7 to $20 per month, and genealogy bundles that include restoration features (MyHeritage, Ancestry) often price at $99 to $199 per year. One-time-purchase desktop options like ON1 Photo Raw 2026.4 cost $99 to $199 with no subscription. Adobe Photoshop's Generative Fill, which added stronger restoration capabilities in 2024, is $23 per month standalone or bundled in Creative Cloud.

The honest value question is: how many photos do you actually need restored? If the answer is 1-5, pay a retoucher. If it is 20-100, a single annual subscription pays for itself within a week of use. If it is in the thousands, look for batch-processing tools with no per-image fees.

Future Developments to Watch

Two trends are worth tracking through 2026. First, on-device AI is becoming capable enough that privacy-sensitive users can restore photos without uploading them to a cloud service; iOS 18 introduced these capabilities in the Photos app, and iOS 26 expanded them with AI Actions for chained commands. Second, video restoration is catching up to photo restoration. Models trained on historical footage can now upscale and colorize silent film with reasonable quality, though temporal consistency between frames remains imperfect.

A third trend worth noting: the restoration market is fragmenting. Niche tools now focus on specific eras (mid-century American), specific damage types (water damage), or specific output goals (printing on canvas). Expect more specialization, not consolidation, through the rest of 2026.

Final Recommendation

For most people in September 2026, the right approach is a hybrid one: scan at high resolution, run the image through a mainstream AI tool like Remini, PicWish, or MyHeritage for the heavy lifting, then spend a few minutes cleaning up artifacts manually in Photoshop or Affinity. Budget between $0 and $20 per month for software, or pay a retoucher $50 to $300 for a small batch of irreplaceable family photos. Do not trust any tool that promises perfect, automated restoration, and always keep an untouched backup of the original scan.