OmniHuman

lastmod 2025-02-13created 2025-02-132 edits

OmniHuman-1: Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Models

A model that generates photorealistic video from a single photo

Features

ByteDance researchers developed an AI system that "generates video of a person naturally talking, singing, and moving from a single photo"
Goes beyond the limitation of earlier AI models, which could only animate the face or upper body, and can generate "video reflecting full-body motion"
The first unified model that realistically animates human motion "using audio, video, or a combination of both as input"
Key features
- Video generation from a single image
  - Can generate natural video reflecting full-body motion from just one photo of a person
- Multimodal input support
  - Supports audio, video, or combined input
  - Greatly improved gesture expressiveness compared to existing models
- Handles images of any aspect ratio
  - Supports various image ratios such as portrait, half-body, and full-body photos
- Compatible with various styles and input data
  - Can reflect various styles such as cartoons, artificial characters, animals, and complex poses
- Natural body movement matched to music style
  - Can generate movement matched to high notes, low notes, and various music genres
- Motion imitation from video
  - Supports video driving, which reproduces a specific person's movements as they are

author Zer0 sitemap