Skip to content

Personalized Restoration via Dual-Pivot Tuning

Personalise image restoration while keeping facial identity

Visit official site

What Personalized Restoration via Dual-Pivot Tuning is

The method is described as a personalised image restoration approach named dual-pivot tuning, built from two steps. It produces natural images that keep personalised facial features together with the degradation attributes of the input.

What you can do with it

  • Personalise a diffusion prior with reference images
  • Keep facial identity inside the denoising framework
  • Retain the visual attributes of degraded images

Application scenarios

  • Restore portraits from a few reference photos of one person
  • Use different personalised models for a face-swap effect
  • Combine text prompts with a personalised model for editing

Method notes

  • Step one fine-tunes a conditional generative model to use encoder condition information
  • Step two fixes the generative model and tunes encoder parameters
  • The paper states experiments show higher-fidelity facial images than non-personalized methods

Pros & cons

✓ What we like

  • Both facial features and degradation attributes are kept
  • Text prompts can guide edits on a personalised model

! What to watch out for

  • Personalisation depends on the reference images provided
  • The workflow runs through two tuning steps

FAQ

How many steps does it use?

The material describes two steps: fine-tuning a conditional generative model, then tuning the encoder.

What does the output keep?

The introduction names personalised facial features and image degradation attributes.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review