FlashMotion / README.md
nielsr's picture
nielsr HF Staff
Add model card for FlashMotion
bcf4c0b verified
|
Raw
History Blame
2.49 kB
metadata
pipeline_tag: image-to-video
license: other

FlashMotion: Few-Step Controllable Video Generation with Trajectory Guidance

FlashMotion is a novel training framework designed for few-step trajectory-controllable video generation. It enables precise motion control along predefined trajectories while significantly reducing the computational overhead and time redundancy typically associated with multi-step denoising processes.

Project Page | Paper | GitHub | FlashBench

Abstract

Recent advances in trajectory-controllable video generation have achieved remarkable progress. However, existing methods rely on multi-step denoising, leading to substantial computational overhead. FlashMotion bridges this gap by introducing a three-stage training framework: training a trajectory adapter on a multi-step generator, distilling the generator into a few-step version (FastGenerator), and finally aligning the adapter with the few-step generator using a hybrid diffusion and adversarial objective.

Installation

# Clone this repository.
git clone https://github.com/quanhaol/FlashMotion
cd FlashMotion

# Install requirements
conda create -n flashmotion python=3.10 -y
conda activate flashmotion
pip install -r requirements.txt
pip install flash-attn --no-build-isolation
python setup.py develop

Sample Usage

FlashMotion supports trajectory-controllable video generation using two types of adapters: ResNet and ControlNet. You can run the provided demo scripts for inference:

# Inference using the ControlNet FastAdapter
bash running_scripts/inference/i2v_control_fewstep_controlnet.sh

# Inference using the ResNet FastAdapter
bash running_scripts/inference/i2v_control_fewstep_resnet.sh

You can customize the generation by modifying the --prompt, --image, and --trajectory arguments within the scripts.

Citation

@misc{li2026flashmotionfewstepcontrollablevideo,
      title={FlashMotion: Few-Step Controllable Video Generation with Trajectory Guidance}, 
      author={Quanhao Li and Zhen Xing and Rui Wang and Haidong Cao and Qi Dai and Daoguo Dong and Zuxuan Wu},
      year={2026},
      eprint={2603.12146},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.12146}, 
}