Adaptive Video Distillation

Mitigating Oversaturation and Temporal Collapse in Few-Step Generation

Project Page

Adaptive Video Distillation
Yuyang You*, Yongzhi Li*, Jiahui Li, Yadong Mu, Quan Chen, Peng ...
CVPR 2026

Overview

This is the official repository for ADV (Adaptive Video Distillation) — a video model distillation method based on DMD（Distribution Matching Distillation）. It addresses oversaturation and slow-motion issues in video generation model distillation, and is capable of learning from new data during distillation training.

Environment Setup

conda create -n AVD python=3.10 -y
conda activate AVD
pip install torch torchvision 
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.

Inference Script

python ./tests/wan/test_bidirectional_fewstep.py

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

Please first modify the wandb account information in the respective config.

Bidirectional DMD Training

torchrun --nnodes 1 --nproc_per_node=8 --master_port 29502 \
    causvid/train_distillation_regression.py \
    --config_path configs/wan_bidirectional_dmd.yaml

Citation

Here is a arxiv version citation bib：

@misc{you2026adaptivevideodistillationmitigating,
      title={Adaptive Video Distillation: Mitigating Oversaturation
             and Temporal Collapse in Few-Step Generation},
      author={Yuyang You and Yongzhi Li and Jiahui Li
              and Yadong Mu and Quan Chen and Peng Jiang},
      year={2026},
      eprint={2603.21864},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.21864},
}

Acknowledgments

Our implementation is largely based on the Causvid and Wan model suite.