hugging-apps/next-forcing-world-model

Next Forcing: Causal World Modeling with Multi-Chunk Prediction

This Space demonstrates Next Forcing, a causal autoregressive world model that generates robot manipulation video from initial camera observations and a text instruction.

How it works

1. 1.Provide top-down and wrist camera images (the initial scene).
2. 2.Describe the manipulation task in natural language.
3. 3.The model autoregressively generates future video chunks (4 frames per chunk).

The model is a 5.1B parameter causal video-action transformer (based on the Wan2.1 architecture) that jointly denoises video latents and action latents in a causal, chunk-by-chunk fashion.

Model

• —Checkpoint: `gangweix/next-forcing-base` (5.1B, BF16)
• —Paper: arXiv:2606.11187
• —Code: github.com/gangweix/next-forcing

Reference

@article{xu2026next,
  title={Next Forcing: Causal World Modeling with Multi-Chunk Prediction},
  author={Xu, Gangwei and Zhang, Qihang and Zhou, Jiaming and Zhu, Xing and Shen, Yujun and Yang, Xin and Xu, Yinghao},
  journal={arXiv preprint arXiv:2606.11187},
  year={2026}
}