IConFace

IConFace: Identity-Structure Asymmetric Conditioning for Unified Reference-Aware Face Restoration

Project Page

More visual comparisons and paper materials are available on the project page:

https://cosmicrealm.github.io/IConFace/

This repository branch is a paper-oriented overview for IConFace. It contains the method summary and visual results only.

Code coming soon.

Overview

Blind face restoration is highly ill-posed under severe degradation because identity-critical details may be missing from the degraded input. Same-identity references can reduce this ambiguity, but they are not spatially aligned with the target and may differ in pose, expression, illumination, age, makeup, or local facial state.

IConFace addresses this problem with identity-structure asymmetric conditioning. The degraded image is treated as the spatial structure anchor, while same-identity references are distilled into a reliability-weighted global identity anchor. This design allows one checkpoint to work in both reference-aware and no-reference restoration settings while reducing reference-specific appearance over-transfer.

Teaser

Framework

The framework separates the roles of different observations. The degraded input preserves target layout, pose, expression, and local structure, while reference images provide identity evidence when available. The same checkpoint can fall back to no-reference restoration when references are absent.

Reference-Aware Restoration

Reference-aware restoration aims to recover identity-consistent details from one or multiple same-identity references while preserving the target structure in the degraded input.

Reference-Aware Metrics

No-Reference Restoration

When no reference is available, IConFace removes reference evidence and restores the degraded portrait from the input itself, maintaining competitive blind restoration quality under synthetic and real-world degradations.

No-Reference Metrics

Citation

@misc{niu2026iconface,
  title={IConFace: Identity-Structure Asymmetric Conditioning for Unified Reference-Aware Face Restoration},
  author={Niu, Axi and Zhang, Jinyang and Qing, Senyan},
  year={2026},
  eprint={2605.02814},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  doi={10.48550/arXiv.2605.02814},
  url={https://arxiv.org/abs/2605.02814}
}