GFPGAN vs CodeFormer vs RestoreFormer — which should I use?

GFPGAN gives the most natural, neutral cleanup. CodeFormer reconstructs heavily degraded or AI-mangled faces but can over-smooth at high weight. RestoreFormer best preserves the original identity, which matters for portraits of real people.

What does the CodeFormer weight actually do?

CodeFormer's weight (often called fidelity, 0 to 1) trades quality for fidelity. Low values (0.2–0.4) maximise sharp invented detail; high values (0.7–1.0) stay closer to the input but restore less. There is no single right value — match it to how damaged the face is.

Why does my restored face look pasted on?

That is usually 100% blend strength on a face whose lighting or grain differs from the scene. Lower the visibility/blend to 60–80% so the restored face mixes with the original texture, or run a light upscale pass afterwards.

Should I restore before or after upscaling?

Restore first on the original-resolution face, then upscale the whole image. Restoring a small face and then enlarging it keeps the model working with the detail it was trained on; upscaling first can lock in artifacts the restorer then amplifies.

AI Face Restoration Settings Guide

AI face restoration

Generated and upscaled images often produce mangled faces — smeared eyes, asymmetric features, plastic skin. Face-restoration models repair them. GFPGAN, CodeFormer, and RestoreFormer each take a different approach, and the single most important setting is the fidelity weight that balances invented detail against staying true to the original face.

How it works

These models detect faces, crop them, run a generative restoration network, and paste the result back with a blend:

final_face = restored_face × blend  +  original_face × (1 − blend)

Weight / fidelity controls how aggressively the network rebuilds. Low = more new detail, high = closer to input.
Blend strength controls how visible the restoration is in the final composite — your tool against the “pasted-on” look.

CodeFormer exposes an explicit weight; GFPGAN uses a fixed network with a visibility slider; RestoreFormer prioritises identity by design.

Tips and example settings

Light cleanup of a good face: GFPGAN at ~0.5 visibility — neutral, fast.
Badly degraded or AI-warped face: CodeFormer weight 0.5–0.7, blend 0.8.
Portrait of a real, recognisable person: RestoreFormer, blend 0.7, so the identity is preserved.
Restore, then upscale. Fix the small face first; enlarge the whole frame afterwards so artifacts are not amplified.
In Automatic1111, set the restorer under Settings → Face restoration and the CodeFormer weight slider; in ComfyUI, drop in a FaceRestore node and feed the weight as a float input.