PhysOPSD uses executable code as privilege for video generation
Original title: PhysOPSD: Code as Privilege for Physically Plausible Video Generation
Authors say the method beats supervised fine-tuning and reinforcement learning with 3,000 samples while keeping general video quality.
PhysOPSD applies on-policy self-distillation to video models, using executable simulation code as privileged physical context instead of text prompts or reference videos. PhysOPSD-Engine turns real physical scenarios into verifiable simulation programs. Dual-level privilege and controlled physical variations under shared rollout noise are meant to expose causal dynamic changes. The authors report better physical plausibility across diverse backbones than SFT and RL baselines.
Unconfirmed: Gains over SFT and RL are the authors' experimental claims and are not independently verified here.
- Training uses only 3,000 samples, according to the paper.
- Code instantiates an editable physical world rather than a static video.
- The method targets dynamics and complex interactions, not visual fidelity alone.
- General video generation quality is reported as preserved.
- Products
- PhysOPSDPhysOPSD-Engine
From the original article
1 min read at huggingface.co
While modern video generative models achieve remarkable visual fidelity, they remain noticeably brittle when generating physically plausible dynamics and complex interactions. Targeted post-training is crucial, yet supervised fine-tuning suffers from offline state mismatch, and reinforcement learning is hindered by sparse, ill-defined physical reward functions.
By Harold Haodong Chen, Rui Peng, Lingteng Qiu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.