All Alchemy pages

Face restoration

Apply a restoration model to detected faces.

GFPGAN and CodeFormers estimate facial details. Either can improve some images, but can also alter identity, expression, and texture.Compare the result with the original before keeping it. A face that already looks correct may become worse.
GFPGAN (face fix)
Fast, general-purpose face restoration. A good default; makes only subtle changes to identity.
GFPGAN v1.3 (face fix)
An earlier GFPGAN version, kept for compatibility. Slightly softer results than current GFPGAN.
CodeFormers (face fix)
Stronger restoration for badly damaged or very low-resolution faces. Can alter identity or fine details more than GFPGAN.
RestoreFormer (face fix)
Favours fidelity to the original face over aggressive fixing, a gentler alternative to CodeFormers.

Examples

GFPGAN (face fix) GFPGAN
The image that was submitted 806 × 808 Source What GFPGAN (face fix) returned 806 × 808 Result

The source is a deliberately degraded portrait, the blurred and compressed kind of photograph face restoration was built to repair. A face that is already sharp would show nothing.

CodeFormers (face fix) CodeFormers
The image that was submitted 806 × 808 Source What CodeFormers (face fix) returned 806 × 808 Result

The same degraded portrait as the GFPGAN entry. CodeFormer trades identity for fidelity differently, which is the comparison the shared source is here to make.

GFPGAN v1.3 (face fix) GFPGANv1.3
The image that was submitted 806 × 808 Source What GFPGAN v1.3 (face fix) returned 806 × 808 Result

The predecessor GFPGAN weight, more faithful to the face it was given than the default.

RestoreFormer (face fix) RestoreFormer
The image that was submitted 806 × 808 Source What RestoreFormer (face fix) returned 806 × 808 Result

A transformer-based restorer that leans towards realism rather than towards smoothing.

Open the Alchemy studio