FantasyPortrait: Multi-Character Portrait Animation
FantasyPortrait is a DiT-based framework for multi-character portrait animation with audio-driven talking heads. Enhances expression-augmented diffusion transformers.
FantasyPortrait
Wan2.1Portrait AnimationTalking HeadAudio-DrivenDiffusion transformer framework for multi-character portrait animation with audio-driven talking head generation. Supports single and multi-character scenarios with expression augmentation.
| Developer | Fantasy-AMAP |
| Release Date | 2025-07 |
| Architecture | Wan2.1-based Diffusion Transformer |
| Capabilities | Multi-character portrait animation, audio-driven talking head, expression control |
| Dataset | Multi-Expr Dataset (first multi-portrait facial expression video dataset) |
Overview
FantasyPortrait is a Diffusion Transformer (DiT)-based framework for generating high-fidelity, emotion-rich portrait animations in both single and multi-character scenarios. Built on Wan2.1, it addresses key challenges in portrait animation including precise expression alignment, multi-character consistency, and audio-visual synchronization.
The model accepts audio input to drive facial expressions and lip movements, making it suitable for applications such as sign language interpreters, animated news anchors, and communication aids. The accompanying Multi-Expr Dataset provides the first multi-portrait facial expression video dataset for training and evaluation.
ComfyUI Integration
FantasyPortrait is available in ComfyUI via the Kijai WanVideoWrapper nodes. The Wan2_1_FantasyPortrait_fp16.safetensors model file and its fp8 scaled variant are hosted on Hugging Face under Kijai/WanVideo_comfy.
Resources
| Resource | Link |
|---|---|
| Paper | arXiv:2507.12956 |
| GitHub | github.com/Fantasy-AMAP/fantasy-portrait |
| HuggingFace (Kijai) | Kijai/WanVideo_comfy |
| Project Page | fantasy-amap.github.io/fantasy-portrait/ |
Guides and workflows related to this model series.