M.S. candidate · Electronic Information · Tsinghua University

Multimodal video generation · diffusion & flow · post-training

🏛️ Computer Application Technology Laboratory · 👨‍🏫 Advisor: Prof. Chun Yuan

Multimodal video AIGC — text-/image-to-video, diffusion and flow matching, autoregressive T2V / TI2V, long-horizon memory and consistency, physics-aware evaluation.
Post-training (SFT, GRPO, on-policy distillation) with structured credit assignment for video and multi-step agents.

Python · PyTorch · verl · vLLM · Wan2.2 / DanceGRPO · LaTeX

🎬 Event-Span GRPO: Event-level Credit Assignment for Physics-aligned Video Generation
Feiyang Liu (first author)

Event-level group-relative advantages for physics-aligned video RL; DanceGRPO with a Wan2.2-class TI2V backbone.

🎞️ Late Forcing: Low-Noise Memory Injection for Frozen Few-Step Video Generators
Feiyang Liu (first author)

Training-free scheduling of far-memory injection along the denoising trajectory for few-step autoregressive video.

🧩 HeadMoE: Data-free Sparse MoE Merging for Multi-site Specialized Models
Feiyang Liu (first author)

Sparse MoE merging with a shared trunk and training-free routing without pooling raw data.

🤖 robot_phys_video_gen — physics-anchored robot manipulation video generation (WAM-oriented)

• M.S., Electronic Information · Tsinghua University
• B.S., Computer Science and Technology · Northeastern University

_{⭐️ From xiangfeidexiaoyang}