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

2026-06-23 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, medium

Summary

TIGER is a novel structured tri-prior fusion framework designed for high-quality Face Video Restoration (FVR), addressing common challenges like identity shift, viewpoint-entangled guidance, and perceptual realism. Developed by Wenxue Li, Peng Zhang, Yifei Chen, Fei Wang, Daiguo Zhou, and others, TIGER integrates an Identity Prior by injecting subject-discriminative embeddings into the latent space to anchor identity. It constructs a Geometry Prior by lifting 2D reference cues into a disentangled 3D parameter space for temporally consistent structural guidance. Furthermore, it utilizes a video generation model's Generative Prior via a one-step rectified flow for efficiency and realism. The framework incorporates a progressive three-stage training optimization strategy to refine structural fidelity, textural reconstruction, and distribution-level realism. A new large-scale FVR dataset was also created for robust training and evaluation. Extensive experiments demonstrate that TIGER achieves superior performance in both identity fidelity and temporal stability, delivering a high-quality, efficient and identity-consistent FVR.

Key takeaway

For Machine Learning Engineers developing face video restoration systems, you should consider TIGER's tri-prior fusion approach to overcome identity shifts and temporal inconsistencies. Implementing distinct Identity, Geometry, and Generative Priors, combined with a progressive three-stage training, can significantly enhance fidelity and realism. Focus on disentangling 3D parameters for structural guidance and leveraging one-step rectified flow for efficient, high-quality output in your models.

Key insights

TIGER fuses identity, geometry, and generative priors to restore high-fidelity, identity-consistent face videos efficiently.

Principles

• Identity anchoring prevents shifts.
• 3D geometry ensures temporal consistency.
• Progressive training refines realism.

Method

TIGER establishes Identity, Geometry, and Generative Priors, then fuses them. It uses a progressive three-stage training strategy to optimize structural fidelity, textural reconstruction, and distribution-level realism.

In practice

• Restore degraded face videos.
• Preserve identity in dynamic videos.
• Enhance realism in video generation.

Topics

• Face Video Restoration
• Identity Preservation
• 3D Facial Priors
• Generative Models
• Temporal Consistency
• Rectified Flow

Code references

• fudan-generative-vision/DicFace
• wangzhiyaoo/SVFR

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• DVFace: Spatio-Temporal Dual-Prior Diffusion for Video Face Restoration — DVFace uses a one-step, dual-prior diffusion framework for efficient, high-quality video face restoration.
• IConFace: Identity-Structure Asymmetric Conditioning for Unified Reference-Aware Face Restoration — IConFace unifies reference-aware and no-reference face restoration using asymmetric identity and structure conditioning.
• Stream3D: Sequential Multi-View 3D Generation via Evidential Memory — Stream3D uses an evidential memory and evidence score mechanism to enable consistent 3D generation from sequential multi-view inputs without retraining existing models.
• Understanding Identity Continuity in Thermal Video through Scene-Level Consistency — High-precision trajectory relinking using scene-level consistency effectively recovers identity in thermal video MOT.
• ReAge3D: Re-Aging 3D Faces with View Consistency — ReAge3D uses a diffusion-based, center-out propagation strategy to achieve view-consistent, detailed 3D face re-aging, surpassing prior 3D editing methods.
• Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding — Video generation models inherently learn robust 3D structural priors and physical laws for scene understanding.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.