ARTFEED — Contemporary Art Intelligence

FaithfulFaces: AI Framework for Pose-Faithful Identity in Video Generation

ai-technology · 2026-05-07

Researchers have introduced FaithfulFaces, a novel framework for identity-preserving text-to-video generation (IPT2V) that addresses identity distortion under large pose variations and occlusions. The system uses a pose-shared identity aligner with a dictionary and a pose variation-identity invariance constraint to maintain consistent facial identity across dynamic scenes. By incorporating explicit Euler angle embeddings, it creates a pose-faithful facial prior for robust generation. The work is detailed in a preprint on arXiv (2605.04702) and targets improvements in complex dynamic video creation.

Key facts

  • FaithfulFaces is a pose-faithful facial identity preservation learning framework.
  • It improves identity-preserving text-to-video generation (IPT2V).
  • The framework uses a pose-shared identity aligner.
  • It includes a pose-shared dictionary and pose variation-identity invariance constraint.
  • Explicit Euler angle embeddings are used for global facial pose representation.
  • The system addresses identity distortion under large pose variations and occlusions.
  • The research is published as arXiv preprint 2605.04702.
  • The framework targets complex dynamic scenes in video generation.

Entities

Institutions

  • arXiv

Sources