Video-Mirai

Video-Mirai: Autoregressive Video Diffusion Models Need Foresight

• 📄 Paper: https://huggingface.co/papers/2606.03971 (arXiv)
• 💻 Code: https://github.com/y0urOy/Video-Mirai
• 🌐 Project page: https://y0urOy.github.io/Video-Mirai/

Video-Mirai closes the representation-level planning gap of causal video generators by letting future segments supervise the current causal state, only at training time. At inference the foresight encoder and predictor are discarded; the deployed generator keeps its causal architecture, FLOPs, and KV-cache behavior identical to the baseline.

Files

File	What it is
model.pt	Trained Video-Mirai foresight checkpoint (chunk-wise Causal-Forcing + DMD + foresight loss). Contains the student generator, EMA weights, and trainer state. Pass to inference.py via --checkpoint_path.

Quick start

# 1. Clone the code repo
git clone https://github.com/y0urOy/Video-Mirai.git
cd Video-Mirai
pip install -r requirements.txt

# 2. Download the Wan2.1 backbone (required by the model)
hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir wan_models/Wan2.1-T2V-1.3B
hf download Wan-AI/Wan2.1-T2V-14B  --local-dir wan_models/Wan2.1-T2V-14B

# 3. Download this checkpoint
hf download y0urOy/Video-Mirai model.pt --local-dir checkpoints

# 4. Run inference
CKPT=checkpoints/model.pt bash scripts/inference.sh

The trainer writes checkpoints in this layout, so once downloaded the file is loaded directly by torch.load:

import torch
state_dict = torch.load("checkpoints/model.pt", map_location="cpu")
# Keys: 'generator' (student weights), 'generator_ema' (EMA weights), and optimizer/scheduler state.
# inference.py picks 'generator_ema' when launched with --use_ema, otherwise 'generator'.

See inference.py and scripts/inference.sh for the full inference entry point.

Method (short)

The causal generator rolls out causally under the same mask used at inference. A frozen foresight encoder (the same Wan2.1-14B used by the DMD score teacher) reads the completed rollout including future segments and produces future-aware feature targets. A lightweight predictor maps each causal hidden state to its fused target via a cosine loss. After training, the foresight encoder and predictor are discarded. The deployed generator keeps its causal architecture, FLOPs, and KV-cache behavior unchanged.

See the paper and project page for the full method, ablations, and qualitative results.

License

Apache 2.0. The Wan2.1 backbone weights this model depends on are released by the Wan-AI team under their own license; please review and comply with their terms when redistributing.

Citation

@article{yu2026videomirai,
  title={Video-Mirai: Autoregressive Video Diffusion Models Need Foresight},
  author={Yu, Yonghao and Huang, Lang and Li, Runyi and Wang, Zerun and Yamasaki, Toshihiko},
  journal={arXiv preprint arXiv:2606.03971},
  year={2026}
}