Hangyi Kuang1,*, Chengzhi Wang1,*, Yefei He1, Weijie Wang1, Yuke Zhao1,2, Zhuokun Chen3
Zisu Huang2, Yifan Yang2, Xiaoxuan He1,†, Feng Chen4,†, Bohan Zhuang1
1 Zhejiang University 2 Microsoft Research 3 Monash University 4 Adelaide University
* Equal contribution † Corresponding authors
News · Method · Installation · Models & Data · Quick Start · Training · Evaluation · Citation
History Forcing is a unified on-policy distillation framework for few-step autoregressive video generators. It compares teacher predictions with and without informative history, then adds the weighted residual to the score or velocity target used for distillation. This strengthens historical constraints for consistent long-video generation and plausible out-of-sight scene evolution.
History Forcing introduces:
- History-contrast supervision. The AR teacher predicts the same noisy student chunk under two contexts that differ only in additional history. Their residual strengthens historical constraints independently of the teacher's overall generative supervision.
- Long-video consistency. Earlier chunks of the student rollout provide subject and scene information less affected by accumulated drift. The residual corrects the DMD reference score while retaining the bidirectional teacher's visual-quality supervision.
- Out-of-sight evolution. Fixed-camera teacher rollouts model how a scene evolves while the student looks away. Their history guides the EMA teacher's velocity target for on-policy self-distillation at revisits.
▶ Watch the overview video
- 2026-11-03: 📄 Our paper is now available on arXiv.
- 2026-11-02: 🤗 Model checkpoints and datasets are publicly available.
- 2026-10-30: 🚀 We provide inference, training and evaluation code for long-video generation and out-of-sight world modeling.
Use Linux with an NVIDIA GPU and a CUDA toolkit compatible with CUDA 12.8.
conda create -n history-forcing -c conda-forge python=3.12 ffmpeg -y
conda activate history-forcing
pip install -U 'pip<25.3' setuptools==69.5.1 wheel==0.43.0 \
torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0
pip install -r requirements.txtRun all commands from the repository root. MemoBench also requires the evaluation dependencies.
| Model | Backbone | Checkpoint |
|---|---|---|
| Long-video generator | Wan2.1-T2V-1.3B | 🤗 DMD |
| World model | DreamX-World-5B | 🤗 OPSD |
| Autoregressive teacher | Wan2.1-T2V-1.3B | 🤗 AR teacher |
The released DMD checkpoint was trained with a 1.3B bidirectional teacher and the 10-second AR teacher.
The 🤗 History-Forcing-Data dataset contains prepared training inputs, PlayWorld OOS tasks and MemoBench Real scenes. The commands below save weights to models/, data to dataset/ and generated videos to output/.
Download the backbone and generator:
hf download Wan-AI/Wan2.1-T2V-1.3B \
--local-dir models/Wan2.1-T2V-1.3B
hf download xzxxntxdy/History-Forcing dmd/model.pt \
--local-dir models/History-ForcingGenerate a 60s video:
bash scripts/long_video/inference.sh \
--prompt "A musician strumming the strings of an acoustic guitar, lost in the melody of their song." \
--output-dir output/demo/long_videoSee long-video inference for prompt lists, seeds and generation settings.
Download the backbone, generator and prepared inputs:
hf download Wan-AI/Wan2.2-TI2V-5B \
--local-dir models/DreamX-World-5B
hf download GD-ML/DreamX-World-5B config.json model.safetensors \
--local-dir models/DreamX-World-5B
hf download xzxxntxdy/History-Forcing opsd/model.pt \
--local-dir models/History-Forcing
hf download xzxxntxdy/History-Forcing-Data --repo-type dataset \
--local-dir dataset/History-ForcingGenerate from an initial image, a scene prompt and camera actions:
bash scripts/world_model/run.sh inference \
--checkpoint models/History-Forcing/opsd/model.pt \
--num-samples 1See custom world-model inputs for input formats.
For Weights & Biases logging, run wandb login once before training.
In addition to the long-video backbone above, download the AR teacher, initialization weights and text prompts, then start training:
hf download xzxxntxdy/History-Forcing ar_teacher/model.pt \
--local-dir models/History-Forcing
hf download zhuhz22/Causal-Forcing chunkwise/ar_diffusion.pt chunkwise/causal_ode.pt \
--local-dir models/Causal-Forcing
hf download gdhe17/Self-Forcing vidprom_filtered_extended.txt --local-dir dmd/prompts
bash scripts/long_video/train.shSee DMD training for configurations, AR teacher fine-tuning and distributed training.
Use the world-model assets downloaded above:
bash scripts/world_model/run.sh pack
bash scripts/world_model/run.sh trainSee OPSD training for history guidance and LoRA/FSDP settings.
| Task | Benchmarks | Guide |
|---|---|---|
| Long-video generation | VBench · MovieGenBench / VBench-Long | DMD evaluation |
| Out-of-sight world modeling | PlayWorld OOS · MemoBench Real | OPSD evaluation |
configs/
├── long_video/ Long-video configurations
└── world_model/ World-model configurations
scripts/
├── long_video/ Long-video launchers
└── world_model/ World-model launchers
dmd/ Long-video implementation and benchmarks
opsd/ World-model implementation and benchmarks
assets/ Paper figures
If you find this work useful, please cite our paper.
@article{kuang2026historyforcing,
title={History Forcing: Distilling History-Dependent Continuations for Long-Horizon Video World Models},
author={Kuang, Hangyi and Wang, Chengzhi and He, Yefei and Wang, Weijie and Zhao, Yuke and Chen, Zhuokun and Huang, Zisu and Yang, Yifan and He, Xiaoxuan and Chen, Feng and Zhuang, Bohan},
journal={arXiv preprint arXiv:2611.02782},
year={2026}
}We thank Causal Forcing, Self Forcing, DreamX-World, Wan, MemRoPE, PlayWorld and VBench for their open-source contributions.
See LICENSE. Third-party components retain their respective licenses.