Disen-Forcing: Disentangling Semantic Anchoring from Motion for Autoregressive Video Diffusion
Paper β’ 2611.02822 β’ Published β’ 11
Paper | Project page | Code
Disen-Forcing is an autoregressive text-to-video model distilled from Wan2.1-T2V-1.3B. It generates 832Γ480 videos at 16 fps from 5 s to 1 min in 4 sampling steps per chunk, keeping long-range consistency without the low-dynamic motion that a frame sink tends to cause.
disen_forcing.safetensors: model weights (bf16).git clone https://github.com/Dijkstra14/Disen-Forcing && cd Disen-Forcing
# environment: see the repository README
huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir wan_models/Wan2.1-T2V-1.3B
huggingface-cli download ZY-Dijkstra/Disen-Forcing disen_forcing.safetensors --local-dir checkpoints
python inference.py --config configs/inference.yaml --checkpoint checkpoints/disen_forcing.safetensors \
--prompts prompts/MovieGenVideoBench.txt --extended_prompts prompts/MovieGenVideoBench_extended.txt \
--output_dir videos --num_output_frames 126 --seed 0
--num_output_frames counts latent frames: 21 β 5 s, 126 β 30 s, 252 β 1 min.
@inproceedings{yu2026disenforcing,
title = {Disen-Forcing: Disentangling Semantic Anchoring from Motion for Autoregressive Video Diffusion},
author = {Yu, Zhengyang and Hayakawa, Akio and Ishii, Masato and Yu, Qingtao and Li, Xuesong and Shibuya, Takashi and Mitsufuji, Yuki and Zhang, Jing},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026}
}
This model is distilled from Wan2.1 and initialized from Causal Forcing. The code builds on CausVid, Self-Forcing, Infinite-Forcing, Causal Forcing and Self Gradient Forcing.
Apache-2.0, following Wan2.1.
Base model
Wan-AI/Wan2.1-T2V-1.3B