Old Photo Restoration with AI

An online tool for restoring old photographs right in your browser. The neural network finds the faces in a picture and rebuilds blurred features — eyes, teeth, contours — on worn, faded and low-quality photos. Each face is processed separately and carefully blended back in; the background and composition stay untouched, and the tone of the photo (sepia, black-and-white) is preserved pixel for pixel. If the picture has no faces at all, a whole-photo enhancement is applied instead — a ×4 upscale (Real-ESRGAN). Restoration strength is adjusted with a live slider after processing, and the change map shows exactly where the AI touched the pixels.

How to restore an old photo

1
Open a demo or load a photo

Click one of the examples — the result appears instantly. Or drop your own picture: JPEG, PNG, WebP, BMP, GIF.

2
Adjust the processing

Turn on auto-tone if the photo has faded (the tool will suggest it) and set the restoration strength. Meanwhile the AI model loads in the background.

3
Click "Restore"

The neural network finds the faces and restores each one separately. Progress and timing stay on screen, and you can cancel at any point.

4
Compare and download

Drag the before/after slider, inspect faces up close, switch off any bad ones, check the change map — then download PNG, JPEG or WebP.

🔥 Trending right now
🗣️ Text to Speech ✨ Video Enhancer ✂️ Trim Audio ✨ Audio Enhancer 🎭 Voice Changer 🎙️ Vocal Remover 🎬 Create Video from Audio 🎬 Trim Video

Bring faces on vintage snapshots back into focus

The neural network cannot run on this computer The processor lacks WebAssembly SIMD support — a newer browser will not change that. Faces can be restored with this same tool on another device: a newer laptop or a phone. Noise, grain and fading, though, can be fixed right now in Photo Repair — it runs on any computer, and your file travels there without re-uploading.
Remove noise and grain → Photo Repair
Drop image here
JPG, PNG, WEBP, GIF, BMP, HEIC, TIFF, AVIF, HEIF ·
See it on an example — instant, no file needed
Examples: photographs by Lewis Hine, Library of Congress (public domain).
Before After
Demo example
…
— ·— ·—
This photo has faded — auto-tone will bring back contrast and depth.
Faces — click to turn restoration on or off for each one
Restoration strength 80%
Marks show what will be filled from the surrounding pixels: a circle for a speck, a box for a long scratch inside it. Click a mark to keep that place as it is; the side beyond the divider shows how it will look. Eyes, nose and mouth of detected faces are protected automatically; check moles and buttons.
Long scratches and creases are found by a separate neural network — a 75 MB model, downloaded once.
⋯Checking browser capabilities…
AI model unavailable The model file is not responding. This is usually temporary — reload the page in a couple of minutes. If it keeps happening, tell us via "Found a bug?".
This is a HEIC photo from an iPhone Browsers cannot read HEIC directly. Convert the photo to JPG and load it again. HEIC to JPG converter ↗
Face restoration — the RestoreFormer++ model (Apache-2.0), face detection — face-api.js, upscaling — Real-ESRGAN. The AI reconstructs plausible detail: the result is an estimate, not an exact copy of the original face.
Don't know what to try?
800+ tools — open a surprise
🎲 Surprise me

Features

One-click before/after examples — before you even pick a photo
Two real archival photos from 1910–1911 with precomputed results: drag the comparison slider and inspect the restored faces up close.
Identity-preserving face restoration — RestoreFormer++
Each face is aligned to 512×512, restored on its own and blended back with soft feathering. The photo tone is locked: sepia stays sepia, black-and-white stays black-and-white.
Live "Restoration strength" slider — instant, no recompute
The balance between original texture and AI detail changes in real time after processing: no waiting for another pass.
Every face on its own — toggle them one by one
Below the result you get a chip for every detected face. If the AI handled someone's face poorly, switch off just that one and keep the rest.
Change map — an honest answer to "what did the AI touch"
The highlight shows where pixels changed and where the photo stayed untouched. Faces that are too small are honestly skipped rather than invented.
Auto-tone for faded prints and ×4 upscaling
Auto-tone brings contrast back to faded scans with one switch, and Real-ESRGAN upscales the result 4× on download (requires WebGPU).

Questions and answers

Which neural network does it use?

RestoreFormer++ (Apache-2.0 license) for face restoration, face-api.js for detection and Real-ESRGAN ×4 for optional upscaling. Processing runs in your browser via WebGPU or WebAssembly.

Can any blurry face be restored?

No. The neural network reconstructs plausible detail from what survives in the picture. If a face is almost erased or was heavily smoothed by another AI, there is nothing left to restore and the result will be a guess. It is an estimate, not an exact copy of the original face.

Why were some faces skipped?

Very small faces (roughly under 60 pixels) are skipped on purpose: at that size the network does not restore features, it invents them. We find it more honest to leave them alone than to replace someone's face.

What happens if the photo has no faces?

The picture still gets a result: the tool applies a whole-photo enhancement — a ×4 upscale (Real-ESRGAN, needs WebGPU). For very large photos and browsers without WebGPU, a link to the AI Photo Enhancer & Upscaler appears below the result — the file travels over without re-uploading.

What does the "Restoration strength" slider do?

It blends the restored face with the original: 100% means maximum AI detail, lower values keep more of the native texture and grain. It works instantly after processing — nothing is recomputed.

Will the background or the color of the photo change?

No. The tool only touches the face regions, and the tone is locked: sepia stays sepia, black-and-white stays black-and-white. The change map shows the affected spots; optional auto-tone gently stretches the contrast of the whole photo without shifting its hue.

Do I need WebGPU?

No — without WebGPU restoration runs on the CPU (a few seconds per face). WebGPU speeds things up and is only required for the ×4 upscale on download.

There are several faces and only some came out well. Why?

Because each face is handled on its own. The network finds every face in the picture, aligns it to 512×512, restores it separately and blends it back with a soft edge — so what a face gets depends entirely on how much of it was there to begin with. A person in the front row, facing the camera, arrives at that 512×512 with plenty to work from; someone at the back or turned away is being upscaled from a handful of pixels, and there the model has to invent more than it recovers. Use the strength slider to pull those faces back toward the original if the invention looks wrong.

💡 Want us to improve this tool just for you?

We can! Just send us a quick message with your idea. If you'd like to discuss it in detail, leave your email and we'll get back to you. You can stay anonymous.

You May Also Like

🩺 File diagnosis 🩹 Photo Repair 🎨 Photo Colorizer ✨ AI Photo Enhancer & Upscaler 🎛️ Photo editor 🗓 Takeout photo dates ⬛ Redact Photo 🖍️ Photo markup

How do you rate this tool?

5 (1 rating)
Thank you for your rating!
Want to share more? Leave a comment!
Thank you! Your comment will appear after moderation.
Who is this tool for?
🎬 Video Creators ✨ Content Creators 👨‍👩‍👧 Parents & Families 📷 Photographers 🩹 File recovery

This result was generated by a neural network. Any rights that could arise on our side pass to you — perpetually, with no usage fees and no limits on territory, print run or manner of use, commercial use included. The rights are non-exclusive: another user may obtain a similar result. Crediting Tembrica or linking back is not required. The file carries a machine-readable AI-origin marker in its metadata. More in the Terms of Service

Published Updated Author: