LPM-1.0: Video-based Character Performance Model
- URL: https://large-performance-model.github.io/
- Date Saved: 2026-06-21
- Source: GitHub Pages (research project page)
- Tags: creative-ai, ai-tools
- Paper: https://arxiv.org/abs/2504.17825
- Code: https://github.com/Qingzheng-Wang/LPM
- Demo: https://huggingface.co/spaces/AIMaster7/LPM
Authors
Qingzheng Wang*, Zhichao Liao*, Siyu Zhu† — Fudan University
Summary
LPM-1.0 is a universal character performance transfer model. Given a single reference image + a driving video, it animates any character type (realistic human full/half-body, humanoid, cartoon) with facial expressions, hand gestures, and body movements.
Key Innovation: Pose-Free
Uses a lightweight reenactment model to generate driving signals without explicit pose estimation — avoids the brittleness of skeleton-based approaches.
Architecture
- Lightweight Reenactment Model — produces driving signals without pose estimation
- Video Generation Backbone + Reference Attention — pretrained video diffusion model with a custom attention mechanism balancing appearance fidelity vs. motion transfer accuracy
Training: 3-Stage Progressive
- Stage 1: Single-frame self-reenactment
- Stage 2: Multi-frame self-reenactment
- Stage 3: Cross-identity fine-tuning
Results (vs. SOTA)
| Method | FID↓ | FVD↓ | SSIM↑ | PSNR↑ | Expression↑ | Pose↑ |
|---|---|---|---|---|---|---|
| MagicAnimate | 152.3 | 487.6 | 0.621 | 18.92 | 0.412 | 0.534 |
| Animate Anyone | 138.7 | 425.3 | 0.654 | 19.87 | 0.456 | 0.578 |
| UniAnimate | 125.4 | 398.2 | 0.678 | 20.45 | 0.489 | 0.612 |
| MimicMotion | 118.9 | 376.5 | 0.695 | 21.12 | 0.523 | 0.645 |
| LPM-1.0 | 95.2 | 312.8 | 0.742 | 22.86 | 0.587 | 0.698 |
Cross-type transfer (Human→Cartoon, Cartoon→Human, Human→Humanoid) also significantly outperforms all baselines.