Skip to content

CodeFormer

A research project for restoring degraded face images

Visit official site

What CodeFormer is

CodeFormer is presented as a Transformer-based prediction network aimed at restoring faces in mosaicked or otherwise degraded images. The central idea described on the project page is to learn a discrete codebook together with a decoder, which narrows the range of plausible reconstructions and makes the restoration mapping less uncertain. The authors report strong robustness to degradation and say the method works on synthetic datasets as well as real photographs.

What you can do with it

  • Restore high-quality faces from degraded inputs
  • Model global face composition as part of reconstruction
  • Apply a controllable feature transformation

What to watch out for

  • This is a research project page rather than a hosted product
  • Robustness is reported on both synthetic and real datasets

Pros & cons

✓ What we like

  • The architecture and the restoration approach are described on a public project page
  • The authors explicitly discuss both synthetic and real datasets
  • A controllable feature transformation is listed as part of the method

! What to watch out for

  • No hosted service, pricing or API quota is published in the source material
  • No consumer workflow or step-by-step usage instructions are given
  • Rights and consent handling for restored portraits is not described

FAQ

What problem does CodeFormer address?

It is described as a network for restoring faces in degraded or mosaicked images.

What is the core technique?

A learned discrete codebook and decoder are used to reduce uncertainty in the restoration mapping.

Is there a hosted product?

The source material is a research project page and does not describe a hosted service.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review