!!! Check out our new CVPR 2024 paper and code designed for text-conditioned image-to-video generation

LFDM

The pytorch implementation of our CVPR 2023 paper Conditional Image-to-Video Generation with Latent Flow Diffusion Models.

Updates

[Updated on 07/08/2023] Added multi-GPU training codes for MHAD dataset.

[Updated on 05/12/2023] Released a testing demo for NATOPS dataset.

[Updated on 03/31/2023] Added the illustration of training a LFDM for NATOPS dataset.

[Updated on 03/27/2023] Added the illustration of training a LFDM for MHAD dataset.

[Updated on 03/27/2023] Released a testing demo for MHAD dataset.

[Updated on 03/26/2023] Added the illustration of training a LFDM for MUG dataset.

[Updated on 03/26/2023] Now our paper is available on arXiv.

[Updated on 03/20/2023] Released a testing demo for MUG dataset.

Example Videos

All the subjects of the following videos are unseen during the training.

Some generated video results on MUG dataset.

Some generated video results on MHAD dataset.

Some generated video results on NATOPS dataset.

Applied LFDM trained on MUG to FaceForensics dataset.

Pretrained Models

Dataset	Model	Frame Sampling	Link (Google Drive)
MUG	LFAE	-	https://drive.google.com/file/d/1dRn1wl5TUaZJiiDpIQADt1JJ0_q36MVG/view?usp=share_link
MUG	DM	very_random	https://drive.google.com/file/d/1lPVIT_cXXeOVogKLhD9fAT4k1Brd_HHn/view?usp=share_link
MHAD	LFAE	-	https://drive.google.com/file/d/1AVtpKbzqsXdIK-_vHUuQQIGx6Wa5PxS0/view?usp=share_link
MHAD	DM	random	https://drive.google.com/file/d/1BoFPQAeOuHE5wt7h-chhYAO-dU0B1p2y/view?usp=share_link
NATOPS	LFAE	-	https://drive.google.com/file/d/10iyzoYqSwzQ3fZgb6oh3Uay-P7k2A12s/view?usp=share_link
NATOPS	DM	random	https://drive.google.com/file/d/1lSLSzS_KyGvJ7dW3l5hLJLR9k2k8LoU3/view?usp=share_link

Demo

MUG Dataset

Install required dependencies. Here we use Python 3.7.10 and Pytorch 1.12.1, etc.
Run python -u demo/demo_mug.py to generate the example videos. Please set the paths in the code files and config file config/mug128.yaml if needed. The pretrained models for MUG dataset have released.

MHAD Dataset

Install required dependencies. Here we use Python 3.7.10 and Pytorch 1.12.1, etc.
Run python -u demo/demo_mhad.py to generate the example videos. Please set the paths in the code files and config file config/mhad128.yaml if needed. The pretrained models for MHAD dataset have released.

NATOPS Dataset

Install required dependencies. Here we use Python 3.7.10 and Pytorch 1.12.1, etc.
Run python -u demo/demo_natops.py to generate the example videos. Please set the paths in the code files and config file config/natops128.yaml if needed. The pretrained models for NATOPS dataset have released.

Training LFDM

The training of our LFDM includes two stages: 1. train a latent flow autoencoder (LFAE) in an unsupervised fashion. To accelerate the training, we initialize LFAE with the pretrained models provided by MRAA, which can be found in their github; 2. train a diffusion model (DM) on the latent space of LFAE.

MUG Dataset

Download MUG dataset from their website.
Install required dependencies. Here we use Python 3.7.10 and Pytorch 1.12.1, etc.
Split the train/test set. You may use the same split as ours, which can be found in preprocessing/preprocess_MUG.py.
Run python -u LFAE/run_mug.py to train the LFAE. Please set the paths and config file config/mug128.yaml if needed.
Once LFAE is trained, you may measure its self-reconstruction performance by running python -u LFAE/test_flowautoenc_mug.py.
Run python -u DM/train_video_flow_diffusion_mug.py to train the DM. Please set the paths and config file config/mug128.yaml if needed.
Once DM is trained, you may test its generation performance by running python -u DM/test_video_flow_diffusion_mug.py.

MHAD Dataset

Download MHAD dataset from their website.
Install required dependencies. Here we use Python 3.7.10 and Pytorch 1.12.1, etc.
Crop the video frames and split the train/test set. You may use the same cropping method and split as ours, which can be found in preprocessing/preprocess_MHAD.py.
Run python -u LFAE/run_mhad.py to train the LFAE. Please set the paths and config file config/mhad128.yaml if needed.
Once LFAE is trained, you may measure its self-reconstruction performance by running python -u LFAE/test_flowautoenc_mhad.py.
Run python -u DM/train_video_flow_diffusion_mhad.py to train the DM. Please set the paths and config file config/mhad128.yaml if needed.
Once DM is trained, you may test its generation performance by running python -u DM/test_video_flow_diffusion_mhad.py.

NATOPS Dataset

Download NATOPS dataset from their website.
Install required dependencies. Here we use Python 3.7.10 and Pytorch 1.12.1, etc.
Segment the video and split the train/test set. You may use the same segmenting method and split as ours, which can be found in preprocessing/preprocess_NATOPS.py.
Run python -u LFAE/run_natops.py to train the LFAE. Please set the paths and config file config/natops128.yaml if needed.
Once LFAE is trained, you may measure its self-reconstruction performance by running python -u LFAE/test_flowautoenc_natops.py.
Run python -u DM/train_video_flow_diffusion_natops.py to train the DM. Please set the paths and config file config/natops128.yaml if needed.
Once DM is trained, you may test its generation performance by running python -u DM/test_video_flow_diffusion_natops.py.

Citing LFDM

If you find our approaches useful in your research, please consider citing:

@inproceedings{ni2023conditional,
  title={Conditional Image-to-Video Generation with Latent Flow Diffusion Models},
  author={Ni, Haomiao and Shi, Changhao and Li, Kai and Huang, Sharon X and Min, Martin Renqiang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={18444--18455},
  year={2023}
}

For questions with the code, please feel free to open an issue or contact me: homerhm.ni@gmail.com

Acknowledgement

Part of our code was borrowed from MRAA, VDM, and LDM. We thank the authors of these repositories for their valuable implementations.

nihaomiao/CVPR23_LFDM