From Slow Bidirectional to Fast Autoregressive Video Diffusion Models [Huggingface][Project]

Few-step Text-to-Video Generation.

From Slow Bidirectional to Fast Autoregressive Video Diffusion Models,
Tianwei Yin*, Qiang Zhang*, Richard Zhang, William T. Freeman, Frédo Durand, Eli Shechtman, Xun Huang (* equal contribution)
CVPR 2025 (arXiv 2412.07772)

Abstract

Current video diffusion models achieve impressive generation quality but struggle in interactive applications due to bidirectional attention dependencies. The generation of a single frame requires the model to process the entire sequence, including the future. We address this limitation by adapting a pretrained bidirectional diffusion transformer to an autoregressive transformer that generates frames on-the-fly. To further reduce latency, we extend distribution matching distillation (DMD) to videos, distilling 50-step diffusion model into a 4-step generator. To enable stable and high-quality distillation, we introduce a student initialization scheme based on teacher's ODE trajectories, as well as an asymmetric distillation strategy that supervises a causal student model with a bidirectional teacher. This approach effectively mitigates error accumulation in autoregressive generation, allowing long-duration video synthesis despite training on short clips. Our model achieves a total score of 84.27 on the VBench-Long benchmark, surpassing all previous video generation models. It enables fast streaming generation of high-quality videos at 9.4 FPS on a single GPU thanks to KV caching. Our approach also enables streaming video-to-video translation, image-to-video, and dynamic prompting in a zero-shot manner.

⚠️ This repo is a work in progress. Expect frequent updates in the coming weeks.

Environment Setup

conda create -n causvid python=3.10 -y
conda activate causvid 
pip install -r requirements.txt 
python setup.py develop

Also download the Wan base models from here and save it to wan_models/Wan2.1-T2V-1.3B/

Inference Example

First download the checkpoints: Autoregressive Model, Bidirectional Model 1 or Bidirectional Model 2 (performs slightly better).

Autoregressive 3-step 5-second Video Generation

python minimal_inference/autoregressive_inference.py --config_path configs/wan_causal_dmd.yaml --checkpoint_folder XXX  --output_folder XXX   --prompt_file_path XXX 

Autoregressive 3-step long Video Generation

python minimal_inference/longvideo_autoregressive_inference.py --config_path configs/wan_causal_dmd.yaml --checkpoint_folder XXX  --output_folder XXX   --prompt_file_path XXX   --num_rollout XXX 

Bidirectional 3-step 5-second Video Generation

python minimal_inference/bidirectional_inference.py --config_path configs/wan_bidirectional_dmd_from_scratch.yaml --checkpoint_folder XXX  --output_folder XXX   --prompt_file_path XXX 

Training and Evaluation

Dataset Preparation

We use the MixKit Dataset (6K videos) as a toy example for distillation.

To prepare the dataset, follow these steps. You can also download the final LMDB dataset from here

# download and extract video from the Mixkit dataset 
python distillation_data/download_mixkit.py  --local_dir XXX 

# convert the video to 480x832x81 
python distillation_data/process_mixkit.py --input_dir XXX  --output_dir XXX --width 832   --height 480  --fps 16 

# precompute the vae latent 
torchrun --nproc_per_node 8 distillation_data/compute_vae_latent.py --input_video_folder XXX  --output_latent_folder XXX   --info_path sample_dataset/video_mixkit_6484_caption.json

# combined everything into a lmdb dataset 
python causvid/ode_data/create_lmdb_iterative.py   --data_path XXX  --lmdb_path XXX

Training

Bidirectional DMD Training

torchrun --nnodes 8 --nproc_per_node=8 --rdzv_id=5235 \
    --rdzv_backend=c10d \
    --rdzv_endpoint $MASTER_ADDR causvid/train_distillation.py \
    --config_path  configs/wan_bidirectional_dmd_from_scratch.yaml  --no_visualize

ODE Dataset Generation. We generate a total of 1.5K dataset pairs, which can also be downloaded from here

torchrun --nproc_per_node 8 causvid/models/wan/generate_ode_pairs.py --output_folder  XXX --caption_path sample_dataset/mixkit_prompts.txt

python causvid/ode_data/create_lmdb_iterative.py  --data_path XXX --lmdb_path XXX 

Causal ODE Pretraining. You can also skip this step and download the ode finetuned checkpoint from here

torchrun --nnodes 8 --nproc_per_node=8 --rdzv_id=5235 \
    --rdzv_backend=c10d \
    --rdzv_endpoint $MASTER_ADDR causvid/train_distillation.py \
    --config_path  configs/wan_causal_ode.yaml  --no_visualize

Causal DMD Training.

torchrun --nnodes 8 --nproc_per_node=8 --rdzv_id=5235 \
    --rdzv_backend=c10d \
    --rdzv_endpoint $MASTER_ADDR causvid/train_distillation.py \
    --config_path  configs/wan_causal_ode.yaml  --no_visualize

TODO

• Checkpoints trained on larger / higher quality dataset.
• Image to Video Generation
• Caching of cross-attention features

Notes

• With the toy dataset, the performance saturates around 1K iterations.
• DMD training likely requires larger, higher-quality datasets.
• Timestep shift, guidance scale, or denoising steps may need fine-tuning.

Citation

If you find CausVid useful or relevant to your research, please kindly cite our papers:

@inproceedings{yin2025causvid,
    title={From Slow Bidirectional to Fast Autoregressive Video Diffusion Models},
    author={Yin, Tianwei and Zhang, Qiang and Zhang, Richard and Freeman, William T and Durand, Fredo and Shechtman, Eli and Huang, Xun},
    booktitle={CVPR},
    year={2025}
}

@inproceedings{yin2024improved,
    title={Improved Distribution Matching Distillation for Fast Image Synthesis},
    author={Yin, Tianwei and Gharbi, Micha{\"e}l and Park, Taesung and Zhang, Richard and Shechtman, Eli and Durand, Fredo and Freeman, William T},
    booktitle={NeurIPS},
    year={2024}
}

@inproceedings{yin2024onestep,
    title={One-step Diffusion with Distribution Matching Distillation},
    author={Yin, Tianwei and Gharbi, Micha{\"e}l and Zhang, Richard and Shechtman, Eli and Durand, Fr{\'e}do and Freeman, William T and Park, Taesung},
    booktitle={CVPR},
    year={2024}
}

Acknowledgments

Our implementation is largely based on the Wan model suite.