GitHub Pages (research project page)
creative-aiai-tools
Original source

LPM-1.0: Video-based Character Performance Model

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

  1. Lightweight Reenactment Model — produces driving signals without pose estimation
  2. 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)

MethodFID↓FVD↓SSIM↑PSNR↑Expression↑Pose↑
MagicAnimate152.3487.60.62118.920.4120.534
Animate Anyone138.7425.30.65419.870.4560.578
UniAnimate125.4398.20.67820.450.4890.612
MimicMotion118.9376.50.69521.120.5230.645
LPM-1.095.2312.80.74222.860.5870.698

Cross-type transfer (Human→Cartoon, Cartoon→Human, Human→Humanoid) also significantly outperforms all baselines.