History Forcing is a few-step autoregressive video generation distillation framework proposed by Zhejiang University and others: it compares the teacher model's predictions with and without historical information, using this difference to correct the distillation objective. The authors claim it improves long-video consistency and the evolution of out-of-view scenes. The repository releases the DreamX-World-5B version, the DMD version (including EMA) of Wan2.1-T2V-1.3B, and the autoregressive teacher model. Inference and training code are available on GitHub.
Why It Matters Few-step autoregressive video generation tends to forget earlier frames as generation progresses. This method uses historical information to correct the distillation objective, targeting long-video consistency.
Hugging Face Likes 4 · Hugging Face Trending Score 3 · Hugging Face Downloads 0
License: Commercial use allowed (base model restrictions apply) · ComfyUI: Not available · VRAM: Not specified
Reasons
- Hugging Face text-to-video trending list, Hugging Face world-model tag
- Score 28 (freshness 0 / source 10 / heat 0 / comfy 0 / vram 3 / demo 5 / license 10)
Usage
- Download weights and run inference according to the instructions in the GitHub repository ziplab/History-Forcing
Risks
- Research code, no ComfyUI support, model card does not specify VRAM requirements; repository does not specify a license