Vera: Identity-Faithful Human Subject-to-Video Generation

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Vera is a unified human-centric Subject-to-Video (S2V) framework designed for identity-faithful single- and multi-person video generation. Developed by Kuaishou Technology, it addresses the challenge of preserving identity-critical human details across frames, poses, and interactions, especially in multi-person scenarios where identity confusion can occur. Vera introduces a million-pair identity-aligned human image-video dataset, constructed via person-level cross-clip retrieval, providing explicit identity correspondence. The framework integrates Identity-Focal Masked Supervision (IFMS) for spatially focused, identity-aware learning and Reference-Aware Layer-wise Attention (RALA) to regulate video token interaction with reference identity cues in the DiT backbone. Experiments show Vera achieves FaceSim-Arc of 0.571 and FaceSim-Cur of 0.529, outperforming baselines like VACE-14B (0.531, 0.507) in identity consistency, while also improving naturalness (NaturalScore 3.90) and reducing reference-image copying.

Key takeaway

For Machine Learning Engineers developing human-centric video generation systems, Vera offers a robust approach to overcome identity drift and confusion. You should consider adopting identity-aligned cross-clip datasets and integrating region-aware supervision like IFMS, alongside layer-wise attention mechanisms such as RALA, to significantly improve facial identity consistency and multi-person subject binding in your models. This framework enhances both identity fidelity and motion naturalness, crucial for personalized storytelling and digital human creation.

Key insights

Vera ensures stable human identity in generated videos by combining a unique dataset with focused supervision and attention mechanisms.

Principles

• Cross-clip data prevents shortcut learning.
• Focus supervision on identity-critical regions.
• Protect reference tokens as stable anchors.

Method

Vera constructs a million-pair identity-aligned dataset via cross-clip retrieval. It uses Identity-Focal Masked Supervision (IFMS) for face-focused loss and Reference-Aware Layer-wise Attention (RALA) to manage reference-video interaction in the DiT backbone.

In practice

• Use cross-clip data for robust identity learning.
• Apply region-aware masks for facial detail emphasis.
• Implement layer-selective attention for identity stability.

Topics

• Human-centric Video Generation
• Identity Preservation
• Diffusion Transformers
• Subject-to-Video (S2V)
• Masked Supervision
• Attention Mechanisms
• Dataset Construction

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

Related on AIssential

• Vera: Identity-Faithful Human Subject-to-Video Generation — Identity-faithful human video generation requires explicit identity correspondence and focused attention mechanisms.
• ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation — Dynamic identity memory and counterfactual self-supervision significantly enhance subject-preserving video generation robustness.
• GroupVideo: Multi-Identity Customized Text-to-Video Generation — GroupVideo enables natural, consistent multi-identity text-to-video generation via multimodal identity alignment and spatial localization.
• Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment — Aura unifies multi-subject video generation by VLM-grounded semantic alignment and a robust data pipeline.
• Navigating User Behavior toward Personalized Multimodal Generation — NaviGen translates user history into executable instructions for personalized multimodal generation via dual codes and a two-stage SFT+RL pipeline.
• GenLCA: 3D Diffusion for Full-Body Avatars from In-the-Wild Videos — GenLCA generates photorealistic 3D avatars from partial 2D video data using a novel visibility-aware diffusion training strategy.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.