Reproduce “ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinye Li, Lingshuai Lin, Lei Wang, Liuzhou Zhang, Jialin Cui, Qingshan Li, Guanchu Wang, Qingbin Liu and 3 more
Researchers from The Chinese University of Hong Kong, Fudan University, Tencent PCG, and Shanghai AI Laboratory developed ForgeWM, a progressive four-stage training framework to transform bidirectional video generators into few-step causal world models. This framework enables low-latency, action-conditioned video generation for interactive environments, achieving up to 72.10 FPS with 168.2 ms latency per chunk for 1-step models while maintaining high visual quality and action accuracy, and supporting replay-time refinement.
Key insights
- Progressive training, particularly the on-policy distribution matching stage (Stage 3), is crucial for mitigating exposure bias and achieving stable few-step autoregressive generation.
- Effective integration of game-native controls via a specialized ActionModule is vital for maintaining high action accuracy and control fidelity in generative world models.
- Budget-specialized causal models and replay-time refinement offer practical trade-offs between latency and visual quality, allowing the same low-step model to be used for both initial draft generation and subsequent quality enhancement without extra training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.