Restore faces in photos with AI
Old photos, faces blurred by enlargement, small faces in group shots: GFPGAN running inside your browser redraws them. Faces are detected automatically and several are handled at once. Photos never leave your device.
The first run downloads an AI model (about 165 MB); later runs use the copy stored in your browser. Your files are never uploaded.
Settings
How to use
- 1
Add photos
Drop blurry faces or old photos. Small faces in group photos are handled too.
- 2
Set the strength
1 uses the restored face as-is; lower values blend in the original to keep the person’s likeness. The first run downloads the AI model (about 165 MB). With WebGPU each face takes about a second; without it, 10–20 seconds on the CPU.
- 3
Run and check
Use “Compare” to check against the original. Only faces are changed, so combine with “AI upscale” if the rest needs work.
How to set the restoration strength and why some faces are not restored are explained in the article “How to Restore Faces in Old Photos with AI” on our blog.
How it works and what to expect
RetinaFace locates each face and its eyes, nose and mouth; each face is cut out at 512×512, restored with GFPGAN v1.4 and pasted back. It is the same combination the official tools use to fix the facial detail that Real-ESRGAN (AI upscale) handles poorly.
Because GFPGAN generates a plausible face, the result can look slightly different from the person. For ID photos and other cases where likeness matters, lower the strength to blend in the original. With WebGPU each face takes about a second; without it, 10–20 seconds on the CPU. The model (about 165 MB) is downloaded once.
Frequently asked questions
Are my photos sent to a server?
No. The AI models are downloaded into your browser and both detection and restoration run on your device. Processing statistics used to fix bugs (device specs, error types and so on; never the photos themselves or their file names) are sent to this site's server. You can turn this off with the switch under “Help us improve the service” at the bottom of the page.
Does it improve anything besides faces?
Only the detected face regions are redrawn. To improve the whole photo, run “AI upscale” first and then face restoration.
Why does the face no longer look like the person?
That is a limit of generative models. Lower the strength to 0.5–0.7 to keep more of the original features. Very low-quality inputs rely more on guesswork.
Does it work without WebGPU?
Yes. Without WebGPU it runs on the CPU (WebAssembly) at 10–20 seconds per face. Afterwards you see whether it ran on the GPU or CPU.
Help us improve the service
If a conversion fails, sending us the circumstances helps us track down the cause. Knowing which devices and formats fail most often, and how long processing takes, lets us revise the supported formats and fix the bugs behind them. If you don't mind, leaving this on is a real help.
What is sent
- The tool you used and the settings you chose (output format, quality, etc.)
- Your device specs (GPU, memory, core count)
- The input format with a rough size
- How many files succeeded or failed and how long it took
- The error type and details when something fails
What is never sent
- The files themselves and file names
- Text, subtitles or prompts you add to a video
- The text to be read aloud, voice descriptions, the pronunciation dictionary and saved voices
- The text in PDFs, the text you rewrote or redacted, passwords, font names and the page ranges you enter
- IP addresses
The data goes only to this site's own server, never to a third-party service.
Details sent by specific tools (counts, times, kinds and yes/no only)
In addition to “What is sent”, text to speech and the PDF tools send the following details.
- Text to speech
- The approximate number of characters and sentences, how many parts the text was generated in, and how long the AI model took to load.
- PDF tools (all)
- The page count range, whether editing is restricted, whether a permissions password was entered, and whether the PDF is signed or encrypted.
- PDF text editing and redaction
- The numbers of rewrites, deletions, white-outs and redactions, the number of rewritten characters by source (original font or fallback font, the latter as a range) and the fallback typeface, the kinds of reasons why lines cannot be edited, and the number of pages turned into images by redaction and the reasons.
- PDF page operations and conversion between images and PDF
- The numbers of inputs and outputs, of pages rotated, deleted, duplicated or moved, of images inserted as they are, converted, in HEIC format or with corrected orientation, and of pages rendered at a lower resolution (all as ranges), whether the original was updated in place or a new document was created, and whether the original protection was reapplied, form fields were written into the page and bookmarks were added back.
- PDF compression
- The numbers of inputs and outputs, of images downsized, not downsized and black-and-white, of downsized images whose transparency was also downsized, that stayed lossless or that were converted from CMYK to RGB, and how many times the file was compressed (all as ranges), the kinds of reasons images were not downsized, whether the target size was met and the kind of reason it was not met, the range of how much smaller the file got, whether the original protection was reapplied, and whether the original PDF was returned unchanged.
- PDF cropping and page numbers
- The numbers of inputs and outputs and of pages cropped, numbered or with their margins reduced (all as ranges), how the margins were set, how the pages to crop or number were chosen, the number format and position, the range of the starting number, that the original was updated in place, and whether the original protection was reapplied.