🎮 BiWM: Bidirectional Autoregressive Video World Models

July 24, 2026 · View on GitHub

The first open-source training framework for bidirectional autoregressive video world models.

BiWM turns a pretrained bidirectional video diffusion model into an action-controllable, autoregressive world model in two training stages:

Camera-control fine-tuning with 81 discrete camera actions.
Few-step DMD distillation for chunk-by-chunk autoregressive generation.

Unlike causal-attention pipelines, BiWM keeps full bidirectional attention inside each current chunk and its history. It supports t2v, i2v, and v2v conditioning in one model.

Demo

https://github.com/user-attachments/assets/e0f8de57-bc5e-4377-9db5-1dc581eacf03

Supported backbones

Backbone	Stage 1	Stage 2 DMD
Wan2.1-1.3B	✅	✅
Wan2.2-TI2V-5B	✅	✅
HunyuanVideo-1.5-8B	✅ cam-text + discrete action	✅
LTX-Video 2.3-22B	✅	✅

Setup

Use separate environments for Wan and HY15/LTX23.

Wan2.1 / Wan2.2

conda create -n biwm-wan python=3.10 -y
conda activate biwm-wan
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu124
pip install transformers==4.44.2 diffusers==0.31.0 accelerate==1.13.0 \
  tokenizers==0.19.1 numpy==1.26.4 peft==0.19.1 torchao==0.17.0 \
  easydict decord einops safetensors imageio imageio-ffmpeg opencv-python

HunyuanVideo-1.5 / LTX-2.3

conda create -n biwm-hy15 python=3.10 -y
conda activate biwm-hy15
pip install torch==2.9.1 torchvision==0.24.1 --index-url https://download.pytorch.org/whl/cu128
pip install transformers==4.56.0 diffusers==0.35.0 torchao==0.15.0 \
  decord einops safetensors imageio imageio-ffmpeg opencv-python

HY15 uses Qwen2.5-VL; LTX23 uses Gemma-3. See hunyuan/ENV_SETUP.md for HY15-specific setup and troubleshooting.

Data

Download the released dataset:

hf download shaohao011/BiWM --repo-type dataset --local-dir dataset
tar -xf dataset/videos_syn.tar -C dataset
tar -xf dataset/videos_real.tar -C dataset

Expected layout:

dataset/
├── videos/                 # synthetic clips
├── videos_syn.json
├── video_real/             # real-game clips
└── videos_real.json

Each JSON record contains a scene-only caption and action_frames. Camera action action_label = translation * 9 + rotation, producing 81 combined actions.

Training

All scripts run from any working directory, resolve the repository root automatically, and allow paths to be overridden through environment variables.

Wan2.2

bash scripts/wan22/stage1_pretrain.sh
STAGE1_DIR=./logs/wan22/stage1 bash scripts/wan22/stage2_dmd.sh

Wan2.1 uses the corresponding scripts under scripts/wan21/.

HunyuanVideo-1.5 discrete actions

Place the base weights under ckpts/HunyuanVideo-1.5, then run:

bash scripts/hy15/stage1_action.sh
STAGE1_DIR=./logs/hy15/stage1_action bash scripts/hy15/stage2_dmd_action.sh

The action path injects per-latent-frame discrete actions into AdaLN while keeping captions free of camera descriptions.

LTX-Video 2.3

export LTX2_CKPT=./ckpts/LTX-Video-2.3/ltx-2.3-22b-dev.safetensors
export GEMMA_PATH=./ckpts/LTX-Video-2.3/google/gemma-3-12b-it-qat-q4_0-unquantized
bash scripts/ltx23/stage1.sh
STAGE1_DIR=./logs/ltx23/stage1 bash scripts/ltx23/stage2_dmd.sh

Common overrides: BIWM_VIDEO_DIR, BIWM_CAPTION_JSON, OUTPUT_DIR, PYTHON_BIN, MAX_TRAIN_STEPS, and CHECKPOINTING_STEPS.

Inference

Wan distilled-model inference:

python pipelines/wan/infer_stage2.py \
  --generator_ckpt ./logs/wan22/stage2/checkpoint-XXXX \
  --wan_base ./ckpts/Wan2.2-TI2V-5B \
  --mode t2v \
  --prompt "A character explores a forest" \
  --action_frames 'w-8, right-12, s-6' \
  --output ./outputs/dmd_infer.mp4

Repository layout

Path	Purpose
`pipelines/wan/`	Wan Stage 1, DMD, inference, and compression
`pipelines/hy15/`	HY15 Stage 1 and DMD
`pipelines/ltx23/`	LTX23 Stage 1 and DMD
`pipelines/common/`	Shared DMD, optimizer, and control code
`wan/`, `hunyuan/`, `ltx23/`	Backbone implementations
`scripts/`	Public training and inference entrypoints

Citation

@article{rui2026biwm,
  title={BiWM: Advancing Open-Source Interactive Video World Models with Bidirectional Autoregression},
  author={Rui, Shaohao and Mao, Xiaofeng and Zhang, Zhanyu and Lin, Peijia and Zhu, Yansong and Zhang, Yibo and Wan, Haibin and Ma, Weijie},
  journal={arXiv preprint arXiv:2606.10135},
  year={2026}
}

Acknowledgements

Built upon FastVideo, minWM, Wan, HunyuanVideo-1.5, and LTX-Video.

License

Released under the Apache License 2.0.