IConFace: AI Framework for Reference-Aware Face Restoration
A novel AI framework named IConFace has been unveiled to tackle the issue of face restoration when significant degradation obscures identity-specific features. This framework is outlined in the arXiv paper 2605.02814 and allows for restoration based on up to three reference images of the same individual, aiding in the recovery of consistent localized characteristics. IConFace employs a hybrid concat backbone that maintains both degraded and reference inputs as dense visual tokens, ensuring the retention of localized reference data. It includes an identity pathway for compact multi-reference guidance and a degraded-structure pathway that integrates full-field and local-residual memories to enhance structure alignment. Additionally, the paper presents a human-audited benchmark to evaluate the survival of localized identity details post-restoration. The announcement type is replace-cross, signaling a revision. This research is significant for the AI image restoration and identity preservation sectors.
Key facts
- IConFace is a fine-grained identity-conditioned framework for face restoration.
- It conditions restoration on up to three same-identity references.
- The hybrid concat backbone retains degraded and reference observations as dense visual tokens.
- An identity pathway provides compact multi-reference guidance.
- A degraded-structure pathway injects full-field and local-residual memories.
- A human-audited benchmark measures whether persistent localized identity details survive restoration.
- The paper is arXiv:2605.02814v2 with announcement type replace-cross.
- The framework addresses underdetermined restoration when severe degradation removes person-specific evidence.
Entities
Institutions
- arXiv