1. Installation
2. Dataset preparation
3. Training
4. Testing
5. Acknowledgments
6. Contacts

Abstract: Flow matching is a recent framework to train generative models that exhibits impressive empirical performance while being relatively easier to train compared with diffusion-based models. Despite its advantageous properties, prior methods still face the challenges of expensive computing and a large number of function evaluations of off-the-shelf solvers in the pixel space. Furthermore, although latent-based generative methods have shown great success in recent years, this particular model type remains underexplored in this area. In this work, we propose to apply flow matching in the latent spaces of pretrained autoencoders, which offers improved computational efficiency and scalability for high-resolution image synthesis. This enables flow-matching training on constrained computational resources while maintaining their quality and flexibility. Additionally, our work stands as a pioneering contribution in the integration of various conditions into flow matching for conditional generation tasks, including label-conditioned image generation, image inpainting, and semantic-to-image generation. Through extensive experiments, our approach demonstrates its effectiveness in both quantitative and qualitative results on various datasets, such as CelebA-HQ, FFHQ, LSUN Church & Bedroom, and ImageNet. We also provide a theoretical control of the Wasserstein-2 distance between the reconstructed latent flow distribution and true data distribution, showing it is upper-bounded by the latent flow matching objective.

Details of the model architectures and experimental results can be found in our following paper:

@article{dao2023lfm,
    author    = {Quan Dao and Hao Phung and Binh Nguyen and Anh Tran},
    title     = {Flow Matching in Latent Space},
    journal   = {arXiv preprint arXiv:2307.08698},
    year      = {2023}
}

Please CITE our paper whenever this repository is used to help produce published results or incorporated into other software.

Python 3.10 and Pytorch 1.13.1/2.0.0 are used in this implementation. Please install required libraries:

pip install -r requirements.txt

For CelebA HQ 256, FFHQ 256 and LSUN, please check NVAE's instructions out.

For higher resolution datasets (CelebA HQ 512 & 1024), please refer to WaveDiff's documents.

For ImageNet dataset, please download it directly from the official website.

All training scripts are wrapped in run.sh. Simply comment/uncomment the relevant commands and run bash run.sh.

Run run_test.sh / run_test_cls.sh with corresponding argument's file.

bash run_test.sh <path_to_arg_file>

Only 1 gpu is required.

MODEL_TYPE=DiT-L/2
EPOCH_ID=475
DATASET=celeba_256
EXP=celeb_f8_dit
METHOD=dopri5
STEPS=0
USE_ORIGIN_ADM=False
IMG_SIZE=256

Argument's files and checkpoints are provided below:

Please put downloaded pre-trained models in saved_info/latent_flow/<DATASET>/<EXP> directory where <DATASET> is defined as in bash_scripts/run.sh.

METHOD=heun
STEPS=50

To evaluate FID scores, please download pre-computed stats from here and put it to pytorch_fid.

Then run bash run_test_ddp.sh for unconditional generation and bash run_test_cls_ddp.sh for conditional generation. By default, multi-gpu sampling with 8 GPUs is supported for faster compute.

python pytorch_fid/compute_dataset_stat.py \
  --dataset <dataset> --datadir <path_to_data> \
  --image_size <image_size> --save_path <path_to_save>

Our codes are accumulated from different sources: EDM, DiT, ADM, CD, Flow Matching in 100 LOC by François Rozet, and WaveDiff. We greatly appreciate these publicly available resources for research and development.

If you have any problems, please open an issue in this repository or ping an email to [email protected] / [email protected].