ECCV 2026

TIGER

Taming Identity, Geometry, and Generative Priors for High-Quality Face Video Restoration

Yang Zhou^1,* · Wenxue Li^1,2,* · Peng Zhang^1,† · Yifei Chen¹ · Fei Wang¹ · Daiguo Zhou¹

¹ MiLM Plus, Xiaomi Inc. ² The Hong Kong University of Science and Technology (Guangzhou)

^* Equal contribution ^† Corresponding author

Abstract

Face Video Restoration (FVR) aims to recover high-fidelity facial videos from degraded input while preserving identity and semantic consistency across frames. Existing methods often struggle to simultaneously address identity shift, viewpoint-entangled guidance, and perceptual realism.

We propose TIGER, a structured tri-prior fusion framework that tames Identity, Geometry, and gEnerative pRiors for high-quality FVR. TIGER anchors identity with subject-discriminative embeddings, lifts reference cues into a disentangled 3D geometry prior, and harnesses a video generative prior through one-step rectified flow.

A progressive three-stage training strategy refines structural fidelity, textural reconstruction, and distribution-level realism. Extensive experiments demonstrate state-of-the-art identity fidelity and temporal stability with efficient, identity-consistent face video restoration.

Method

A tri-prior framework for identity-consistent, geometry-aware, and realistic restoration.

Identity Prior

Reference identity embeddings are injected into the latent backbone to anchor subject-discriminative facial traits under severe degradation.

Geometry Prior

2D reference cues are lifted into a 3D parameter space and fused with frame motion to provide temporally consistent structure.

Generative Prior

A one-step rectified-flow formulation transports degraded latents to the high-fidelity manifold while preserving photorealistic detail.

TIGER framework pipeline — TIGER integrates identity, geometry, and generative priors with progressive optimization.

Restoration Demos

Paired degraded inputs, restored outputs, and identity references from VFHQ and VOX2.

VFHQ Sample 01Clip+1qf8dZpLED0

Reference

Degraded Input

TIGER Restoration

VFHQ Sample 02Clip+1L2d-mQA-Gc

Reference

Degraded Input

TIGER Restoration

VOX2 Sample 01id00088

Reference

Degraded Input

TIGER Restoration

VOX2 Sample 02id00109

Reference

Degraded Input

TIGER Restoration

Qualitative Results

Additional comparisons from the paper.

BibTeX

@inproceedings{zhou2026tiger,
  title={TIGER: Taming Identity, Geometry, and Generative Priors for High-Quality Face Video Restoration},
  author={Yang Zhou and Wenxue Li and Peng Zhang and Yifei Chen and Fei Wang and Daiguo Zhou},
  booktitle={European Conference on Computer Vision},
  year={2026}
}