How to Restore Faces with CodeFormer

Face restoration is not ordinary sharpening. CodeFormer reconstructs facial detail, so clearer output may also drift away from the person's real appearance. The key is to begin with high fidelity and evaluate identity, not beauty.

Choose the correct restoration mode

ModeWhat it doesUse it when
Standard face restorationDetects faces, restores them, and pastes them back at the original image sizeFaces are compressed, blurred, or lack detail
White-mask reconstructionReconstructs a marked missing area and produces a 512×512 face resultA face has a localized missing or occluded region

Generated result: reconstruction does not prove what the person actually looked like. Keep and clearly label the original.

Restore a face without losing identity

Use the clearest source

Start with the best available scan and correct rotation and overall brightness before face restoration.

Decide whether to restore only the main face

For a group photo focused on one person, this avoids unnecessary reconstruction of tiny background faces.

Begin with higher fidelity

Higher fidelity usually preserves identity more closely but makes a subtler repair. Save several settings for comparison.

Upscale only when needed

Enable AI upscaling when the final file truly needs larger dimensions, not simply because the option is available.

Verify identity markers

Compare eye shape, eyebrows, mouth, moles, hairline, and face outline against the source.

Prepare a masked reconstruction carefully

Mark only the genuinely missing region with a clean white mask and preserve as much valid facial structure as possible. The model needs surrounding shape, tone, and context. When most identity information is absent, no setting can recreate historical truth.

Use parameters according to purpose

  • For family memorial photos, favor fidelity over dramatic enhancement.
  • For low-resolution web images, reduce fidelity gradually and compare facial features.
  • For group portraits, test the main-face option first.
  • For damaged regions, preserve the mask and multiple candidates.
  • Confirm the face before applying final upscaling or color work.

Frequently asked questions

What fidelity value should I use?

There is no universal value. Start toward source fidelity and adjust gradually while checking whether the person still looks the same.

Why does the restored face look like someone else?

When source detail is missing, the model generates common facial patterns. Use a better scan, increase fidelity, or accept a less aggressive restoration.

Does CodeFormer run on every device?

It requires the appropriate AI runtime and model. Availability and speed depend on the operating system, package, memory, and hardware.