Diffusion Forcing Transformer with History Guidance

Kiwhan Song*¹ · Boyuan Chen*¹ · Max Simchowitz² · Yilun Du³ · Russ Tedrake¹ · Vincent Sitzmann¹
*Equal contribution ¹MIT ²CMU ³Harvard

Paper | Website | HuggingFace Demo | Pretrained Models

This is the official repository for the paper History-Guided Video Diffusion. We introduce the Diffusion Forcing Tranformer (DFoT), a novel video diffusion model that designed to generate videos conditioned on an arbitrary number of context frames. Additionally, we present History Guidance (HG), a family of guidance methods uniquely enabled by DFoT. These methods significantly enhance video generation quality, temporal consistency, and motion dynamics, while also unlocking new capabilities such as compositional video generation and the stable rollout of extremely long videos.

🔥 News

• 2025-02: Diffusion Forcing Transformer is released.

🤗 Try generating videos with DFoT in your browser!

We provide an interactive demo on HuggingFace Spaces, where you can generate videos with DFoT and History Guidance. On the RealEstate10K dataset, you can generate:

• Any Number of Images → Short 2-second Video
• Single Image → Long 10-second Video
• Single Image → Extremely Long Video (like the teaser above!)

Please check it out and have fun generating videos with DFoT!

📖 Detailed Development Wiki

If you just want quickly try our model in python, you can skip this section and proceed to the "Quick Start" section. Otherwise, we provide a comprehensive wiki for developers who want to extend the DFoT framework. It offers a detailed guide on thse topics:

• code structure
• command line options
• datasets structure
• checkpointing and loading
• training and evaluating DFoT or baseline models
• every command to reproduce the paper's result

🚀 Quick Start

Setup

1. Create a conda environment and install dependencies:

conda create python=3.10 -n dfot
conda activate dfot
pip install -r requirements.txt

2. Connect to Weights & Biases:

We use Weights & Biases for logging. Sign up if you don't have an account, and modify wandb.entity in config.yaml to your user/organization name.

Generating Videos with Pretrained Models

Simply run one of the commands below to generate videos with a pretrained DFoT model. They will automatically download a tiny subset of the RealEstate10K dataset and a pretrained DFoT model.

1. Single Image to Long Video (200 Frames, takes ~10 mins):

NOTE: if you encounter CUDA out-of-memory errors (due to limited VRAM), try setting algorithm.tasks.interpolation.max_batch_size=1.

python -m main +name=single_image_to_long dataset=realestate10k_mini algorithm=dfot_video_pose experiment=video_generation @diffusion/continuous load=pretrained:DFoT_RE10K.ckpt 'experiment.tasks=[validation]' experiment.validation.data.shuffle=True dataset.context_length=1 dataset.frame_skip=1 dataset.n_frames=200 algorithm.tasks.prediction.keyframe_density=0.0625 algorithm.tasks.interpolation.max_batch_size=4 experiment.validation.batch_size=1 algorithm.tasks.prediction.history_guidance.name=stabilized_vanilla +algorithm.tasks.prediction.history_guidance.guidance_scale=4.0 +algorithm.tasks.prediction.history_guidance.stabilization_level=0.02  algorithm.tasks.interpolation.history_guidance.name=vanilla +algorithm.tasks.interpolation.history_guidance.guidance_scale=1.5

2. Single Image to Short Video (8 Frames, takes <1 min):

python -m main +name=single_image_to_short dataset=realestate10k_mini algorithm=dfot_video_pose experiment=video_generation @diffusion/continuous load=pretrained:DFoT_RE10K.ckpt 'experiment.tasks=[validation]' experiment.validation.data.shuffle=True dataset.context_length=1 dataset.frame_skip=20 dataset.n_frames=8 experiment.validation.batch_size=1 algorithm.tasks.prediction.history_guidance.name=vanilla +algorithm.tasks.prediction.history_guidance.guidance_scale=4.0

3. Interpolating Two Distant Images into Short Video (8 Frames, takes <1 min):

python -m main +name=two_images_to_interpolated dataset=realestate10k_mini algorithm=dfot_video_pose experiment=video_generation @diffusion/continuous load=pretrained:DFoT_RE10K.ckpt 'experiment.tasks=[validation]' experiment.validation.data.shuffle=True dataset.frame_skip=20 dataset.n_frames=8 experiment.validation.batch_size=1 algorithm.tasks.prediction.enabled=False algorithm.tasks.interpolation.enabled=True algorithm.tasks.interpolation.history_guidance.name=vanilla +algorithm.tasks.interpolation.history_guidance.guidance_scale=4.0

4. Generating Videos on Other Datasets

Please refer to our wiki for more details.

Training

Training a DFoT model requires a large, full dataset. The commands below will automatically download the necessary data, but please note that this process may take a while (~few hours). We also provide specifications for the GPUs required for training. If you are training with fewer GPUs or using a smaller ⁠experiment.training.batch_size, we recommend proportionally reducing ⁠experiment.training.lr. You training will produce a wandb link which ends with a wandb run id. To load or resume your trained model, simply append load={the_wandb_run_id} and resume={the_wandb_run_id} to the training / inferencecommand.

1. RealEstate10K (12 x 80GB VRAM GPUs)

python -m main +name=RE10k dataset=realestate10k algorithm=dfot_video_pose experiment=video_generation @diffusion/continuous

2. Kinetics-600 (12 x 80GB VRAM GPUs)

python -m main +name=K600 dataset=kinetics_600 algorithm=dfot_video experiment=video_generation @DiT/XL

3. Minecraft (12 x 80GB VRAM GPUs)

Note: Minecraft training requires additionally preprocessing videos into latents (see here).

python -m main +name=MCRAFT dataset=minecraft algorithm=dfot_video experiment=video_generation @diffusion/continuous @DiT/B

📝 Acknowledgements

This repo is using Boyuan Chen's research template repo. By its license, we just ask you to keep the above sentence and links in README.md and the LICENSE file to credit the author.

📌 Citation

If our work is useful for your research, please consider giving us a star and citing our paper:

@misc{song2025historyguidedvideodiffusion,
  title={History-Guided Video Diffusion}, 
  author={Kiwhan Song and Boyuan Chen and Max Simchowitz and Yilun Du and Russ Tedrake and Vincent Sitzmann},
  year={2025},
  eprint={2502.06764},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2502.06764}, 
}