Explicit model consent
GFPGAN is requested only after you approve its model download; declining keeps the instant classical restoration path available.
Free local tool
AI face restore
Rebuild detail in a small, soft or damaged face with GFPGAN v1.4, then inspect the result beside the source. The model runs on your device after an explicit one-time download decision; the photograph is never uploaded. A non-generative local fallback is available when you decline or the device cannot run the model.
or drop one here — JPEG, PNG or WebP, up to 32 MB. It never leaves this device.
Before you download GFPGAN v1.4 face restore
Checking what this device can do with GFPGAN v1.4 face restore…
One-time download
GFPGAN v1.4 reconstructs facial detail that is genuinely missing — eyes, teeth and hair in a face that is only a few dozen pixels across. It is a 340 MB one-time download, cached afterwards, and restores the first detected face in roughly seven seconds on a CPU. Crop group photos and run each face separately.
If you decline: a classical bilateral + CLAHE + unsharp pipeline runs instead — instant and no download. It sharpens and evens out what is already there, but it cannot rebuild detail that was never captured.
Choose an image first.
Result proof
A comparison you can audit, not a silent overwrite
01
Input
Untouched portrait
02
Local pass
One face restored locally
03
Result
Separate PNG result
How to verify it: Use the comparison view around eyes, teeth, hairline and jewellery. If identity-bearing detail changed, reduce the strength or keep the original.
One face at a time
Restoration is a proposal you can compare
Face restoration rebuilds plausible facial detail from the pixels that remain. It updates the selected canvas item through normal editor history, so the original and restored versions stay comparable and undoable.
1 selected face
Restoration scope
Canvas result, real history
The restored bitmap replaces the active editor item through the same checkpoint path as other destructive edits.
Identity is not recovered
Fine features are inferred, so compare eyes, teeth, jewellery, and distinguishing marks against the source.
Know before you process: Run one clearly visible face at a time and treat the result as a visual restoration, not evidence of details the camera never captured.
What “restore” honestly means
Lost pixels cannot be recovered. GFPGAN generates a plausible face constrained by what survived in the source. A mildly soft portrait usually retains enough structure to stay recognisable; a severely degraded one gives the model more freedom and may look convincing while changing the person. This is useful restoration for display, not forensic recovery.
What people use this for
Small faces in group photos
Crop one person, restore that face, and compare it at the source scale before placing it back into a larger composition.
Scanned family prints
Improve the face while keeping the untouched scan as the archival record and labelling the restored copy clearly.
Soft social portraits
Recover readable facial structure from a compressed or low-light portrait without sending it to a cloud service.
Restoration before colourising
Repair facial structure first, then colourise. Colour applied to a damaged face makes the later repair harder to judge.
Limits worth knowing
- The current pass restores one face at a time; process group portraits as deliberate crops.
- The model creates plausible detail and can alter identity-bearing features when the source is badly damaged.
- GFPGAN is a large one-time download and may be impractical on low-memory devices; the readiness panel reports the actual path before processing.
- The classical fallback sharpens and balances existing pixels but cannot reconstruct missing facial structure.
How the local workflow runs
- Choose a JPEG, PNG or WebP and inspect the untouched preview.
- Approve or decline the named GFPGAN download; no model file is fetched before that decision.
- Run the restoration and wait for the result label to identify the path that actually ran.
- Compare identity-bearing details at 100% rather than judging only the thumbnail.
- Download a separate PNG and retain the original source file.
Frequently asked questions
Does face restoration recover the real original face?
No. GFPGAN reconstructs a plausible face from the surviving structure. Keep the original and do not use a restored face as evidence or for identification.
Does the photo leave my device?
No. The selected image and result remain in this browser tab. If you approve GFPGAN, only the model files are downloaded to your device.
What happens if I decline the model download?
A clearly labelled classical bilateral, contrast and sharpening pass remains available. It improves existing detail but does not perform generative face reconstruction.
Can it restore a group photograph?
The current pass restores one detected face at a time. Crop and process each person separately, then compare every result against the original group photo.
Why can a severely damaged face look sharp but wrong?
When little identity signal remains, the model prior contributes more of the output. Crispness is not proof of accuracy; it can make uncertainty harder to notice.
Which file should I keep?
Keep the untouched scan as the archival source and save the restoration as a separate PNG with a descriptive filename.
Related tools
Ready to try it?
Private, reversible and free — keep the original beside the result.
Restore a face →