The short answer: the AI of 2026 can sharpen faces, colorize black-and-white photos and even animate your grandmother — but it only works with what you give it. AI software does not restore pixels that were never scanned; it guesses them. Scan quality is therefore the ceiling of everything that follows: printed photos — minimum 600 DPI, slides — 2400 DPI, negatives — 1200 to 2400 DPI. Scan low, and no software on earth will bring back what was left on the scanner glass.
This is the full guide: what DPI is and why it decides everything, what actually happened when researchers fed an AI a photo of Obama, which tools are worth the money, and the right order to do it all. By the end you will know exactly what to do with the box in the attic.
What changed in the AI era — and why scanning right pays off now more than ever
Five years ago a scanned photo was the end of the road: a file, a backup, done. Today it is a starting point. MyHeritage offers automatic colorization and photo animation (LiveMemory); Topaz recovers blurred faces; Photoshop removes scratches in a click. The value of every scanned photo has jumped — but only if the scan contains enough real information. And that is exactly where most people fail: they scan with a phone or a home scanner at low resolution, run AI, and get a result that looks "almost like grandma". Almost.
What is DPI — the explanation anyone can follow
DPI (Dots Per Inch) = how many points of information are sampled from every inch of the original. It is not abstract "quality" — it is the amount of real information pulled from the photo into the file. A 4×6" print scanned at 300 DPI yields 2.2 megapixels; at 600 DPI — 8.6 megapixels, four times as much. With slides it is even more dramatic: a 35mm frame is a tiny 36×24mm piece of film — at 600 DPI you get barely half a megapixel from it. At 2400 DPI — 7.7 megapixels. Sixteen times more.
The dashed line in the chart is the critical point: below roughly 2 megapixels, AI tools do not have enough real structure to work with — and they switch from restoring to inventing.
The experiment that proved it: the white Obama
In 2020, researchers fed a pixelated photo of Barack Obama to an AI upscaling algorithm called PULSE. The result: the face of a white man. Not a bug — structure: infinitely many different faces reduce to the same low-resolution image, and the network simply picks one from what it has learned. LetsEnhance, itself a maker of upscaling software, puts it with rare honesty: "AI doesn't restore lost pixels — it predicts them. When the input is weak, the model fills in details from generic face patterns." In the words of anyone holding a box of family albums: weak scan + AI = a face that is almost grandma, but not grandma.
Four reasons the scan comes before any software
1. Face hallucination. When a face occupies only a few dozen pixels in the scan, the model invents the facial structure — different runs of the same tool produce different faces. 2. Amplified JPEG compression. Upscalers cannot always tell real detail from compression artifacts — they enlarge and sharpen the compression blocks too. A clean scan removes that layer of damage in advance. 3. Colorization needs tonal range. Colorization models map color from grayscale values — a scan with blown highlights or crushed shadows leaves the model no clues, and the result is a uniform sepia smear. 4. Noise versus grain. Only at sufficient resolution can a denoising model tell authentic film grain from real detail — on a low scan it erases both.
How much DPI do you need? The standards — and our recommendations
We did not invent these numbers: the US federal FADGI standard and the Library of Congress guidelines define the professional preservation bar, and our recommendations sit on it:
| Material | Archival standard (FADGI / Library of Congress) | Our recommendation |
|---|---|---|
| Printed photos | 400 DPI, 600 for small prints or enlargement | Minimum 600 DPI — in the AI era, today's small print is tomorrow's enlarged portrait |
| 35mm slides | 2,000–4,000 DPI | 2400 DPI — 7.7MP of real information from every frame |
| Negatives | 2,000–4,000 DPI | 1200–2400 DPI depending on purpose — viewing or reprinting |
Flatbed versus regular scanner — and why the phone is not even in the race
| Method | How it works | What you get |
|---|---|---|
| Phone photo (scanning apps) | The phone camera photographs the print in room light | Convenient — but reflections, perspective distortion and low effective resolution. Exactly the input that limits every AI tool afterwards |
| Feed scanner (regular/office) | Rollers pull the photo across a sensor | Fast for loose photos, but fragile prints are at risk, and there is no transparency light source — slides and negatives are out |
| Professional flatbed | The photo lies still; for transparencies — a dedicated light source above | Zero mechanical risk, true high resolution, and the density range (Dmax) that pulls detail out of shadows — the only way for slides and negatives |
At our lab, loose photos run on the Epson FastFoto FF-640 (one photo per second, a file per photo including backs with dedications), and everything delicate, oversized or transparent — slides, negatives, glued albums — is scanned by hand on the Epson V850 Pro with its dual-lens system.
The recommended AI tools of 2026 — an honest review
| Tool | What it does brilliantly | Price (approx.) | The limitation to know |
|---|---|---|---|
| Topaz Photo AI / Gigapixel | Upscaling, sharpening and face recovery — the professional standard | Subscription only since late 2025, from $149/year | On very small faces it produces "generic" faces; needs a strong computer |
| Remini | Sharpens blurred faces on your phone, impressive in a second | From $7/week (web) | The best-documented criticism: on weak input it invents faces — the "wax" effect |
| MyHeritage | Colorization, scratch repair, animation (LiveMemory) — a moving family package | Free for a few photos; subscription for full use | Colorization is an educated guess — leans to "safe" beige tones |
| Adobe Photoshop | Photo Restoration + Generative Fill — the powerhouse for heavy damage | Creative Cloud subscription | Handles ~70% of damage; a large tear still needs handwork — and Generative Fill invents, not restores |
| Google Photos | Photo Unblur and Magic Eraser — free and accessible | Free (limited to 10 saves/month without a subscription) | No scratch removal or colorization; "storage saver" compression harms the original |
| VanceAI | Automatic scratch and tear removal in the browser | From $9.90/month (credits) | No manual control — on a low scan it "heals" real details too |
| Gemini (Nano Banana) | Restoration and colorization by chat prompt — the big free "wow" of 2026, output up to 4K | Free, ~20 images/day (as of writing) | Repaints the image — an interpretation, not a restoration; invisible digital watermark always |
| ChatGPT (GPT Image 1.5) | Conversational editing and restoration, much-improved face consistency | Free, ~2–3/day | Output capped at 1536px — not for printing; regenerates the whole image |
Notice the pattern in the limitations column: every tool fails in the same place — weak input. None of them compensates for a bad scan; all of them excel on a great one.
What about ChatGPT and Gemini? The tools everyone already has
Most people today do not download software — they upload the photo to a chat and type "restore this". It is worth knowing what actually happens there. Google's Gemini with the Nano Banana image model is the star of the field: the restoration-and-colorization trend that went viral in 2025, and since February 2026 Nano Banana 2 is the default — output up to 4K and roughly 20 free images a day (as of this writing; check the app). ChatGPT with GPT Image 1.5 (December 2025) improved dramatically at preserving facial likeness — an improvement that quietly admits the previous generation's famous problem — but output is capped at 1536 pixels, far from print requirements.
And here is the difference you must understand: dedicated tools like Topaz enhance the existing pixels. Chat models repaint the image — they encode what you uploaded and generate a new image inspired by it. The result is beautiful, sometimes moving to tears — but it is "a new picture of your picture", not the picture itself. In group photos they can blend facial features between people, and ethics research documents a tendency to "correct" faces toward the training-data ideal. The input side is limited too: Gemini downscales any uploaded image to roughly 3,000 pixels at most, ChatGPT to about 2,000 — one more reason the real archive must be the scan file, not the chat.
Our verdict: do it! It is fun, it is accessible, and grandma in color moves the whole family group chat. But always on a copy — with the original full-resolution scan kept aside as your "digital negative". A chat output that replaces the original erases the historical evidence forever. Field tip: add "restore without changing facial features or adding details" to your prompt — it helps, but it is no guarantee.
The right order: scan → original → AI
Step 1: professional scanning at the right DPI (table above). Step 2: keep the original scan file untouched — it is your "digital negative", the file you return to when better software ships in two years. Step 3: only now AI — sharpening, colorizing, animating — always on a copy, never on the original. Step 4: 3-2-1 backup: three copies, two media types, one outside the house. That is why we deliver everything as an organized download link — every album in its own folder, ready for backup and for any AI tool you choose.
And what does it cost with us?
0.79 ILS per loose photo · 0.89 ILS per pocket-album photo · 5.5 ILS per glued-album page (no disassembly) · slides 2.35–3.6 ILS by resolution, including 2400 DPI · negatives 49 ILS per strip. Minimum order 150 ILS, and from 1,000 photos home pickup is free anywhere in Israel. All the numbers and worked examples — in our full photo scanning price guide and slide scanning guide. More than 5,000,000 frames and photos have already passed through our scanners — we know exactly how much information every photo can give.
Frequently asked questions
Can AI software save a photo that was scanned at low quality?
Not really. It will produce a sharp, pretty image — but the added details are the model's guess, not your photo. Especially with faces: the result can be "almost the person, but not them". The right fix is rescanning at high resolution — the physical original still holds the details.
What DPI should old photos be scanned at?
Minimum 600 DPI for printed photos, 2400 DPI for slides, 1200–2400 for negatives. These recommendations align with archival standards (FADGI and the Library of Congress) — and leave AI enough real material to work with.
Which AI tool is best for restoring old photos?
Depends on the task: Topaz Photo AI for professional sharpening and upscaling, MyHeritage for moving colorization and animation, Photoshop for heavy physical damage, Google Photos for quick free fixes. All of them work better — sometimes dramatically — on a high-DPI scan.
Why not just photograph the photos with a phone?
Phone scanning apps are convenient, but add reflections, distortion and low effective resolution — exactly the input that makes AI tools invent details. For a family archive you scan once in a lifetime, the gap between a phone shot and a professional scan is the gap between "I have a copy" and "I have an archive".
Can I just ask ChatGPT or Gemini to restore my photo?
You can, and it is fun — Gemini's Nano Banana gives impressive results for free. Just understand that chat models repaint the image rather than enhance the existing pixels — the result is an interpretation, and faces can subtly change. So always keep the original scan file, and use chat output for sharing and excitement — not as a replacement for your archive.
Do you also do the AI step for me?
We focus on the part that cannot be redone later: extracting the maximum from the original, with the right equipment and hands. With files like that, every tool in this review will work beautifully for you — and for heavy restoration (film, video) we run a professional suite of our own. Tell us what you have and we will point you right.
Want AI working for you instead of against you? It starts with the right scan. Tell us what is in the box or call +972-3-5610368 — and from 1,000 photos, our courier is already on the way. More in English: Studio HanaBi.
