🚀 FlowTurbo: Towards Real-time Flow-Based Image Generation with Velocity Refiner (NeurIPS 2024)

Created by Wenliang Zhao*, Minglei Shi*, Xumin Yu, Jie Zhou, Jiwen Lu

This repo contains PyTorch model definition and training/sampling code for FlowTurbo.

[arXiv]

We propose a framework called FlowTurbo to accelerate the sampling of flow-based models while still enhancing the sampling quality. Our primary observation is that the velocity predictor’s outputs in the flow-based models will become stable during the sampling, enabling the estimation of velocity via a lightweight velocity refiner. Additionally, we introduce several techniques, including a pseudo corrector and sample-aware compilation, to further reduce inference time. Since FlowTurbo does not change the multi-step sampling paradigm, it can be effectively applied to various tasks such as image editing, inpainting, etc. By applying FlowTurbo to different flow-based models, we obtain an acceleration ratio of 53.1%∼58.3% on class-conditional generation and 29.8%∼38.5% on text-to-image generation. Notably, FlowTurbo reaches an FID of 2.11 on ImageNet with 100 (ms / img) and FID of 3.93 with 38 (ms / img), achieving the real-time image generation and establishing the new state-of-the-art.

Setup

First, download and set up the repo:

git clone https://github.com/shiml20/FlowTurbo.git
cd FlowTurbo

We provide an environment.yml file that can be used to create a Conda environment. If you only want to run pre-trained models locally on CPU, you can remove the cudatoolkit and pytorch-cuda requirements from the file.

conda env create -f environment.yml
conda activate FlowTurbo

Sampling

Pre-trained FlowTurbo checkpoints. You can sample from our pre-trained FlowTurbo models with sample.py. Please Download the weights for our pre-trained FlowTurbo model.

python sample.py

For convenience, our pre-trained SiT-XL/2-Refiner models can be downloaded directly here as well:

Training SiT

We provide a training script for FlowTurbo in train.py. To launch FlowTurbo (256x256) training with N GPUs on one node:

CUDA_VISIBLE_DEVICES='0' torchrun --nnodes=1 --nproc_per_node=1 --master_port 12345 train.py \
    --data-path /data/ILSVRC2012/train --global-batch-size 18 \
    --note 'NAME' --ckpt-every 5000 --lr 5e-5 --vae_ckpt vae-ema --model_teacher_ckpt /pretrained_models/predictor.ckpt \

Evaluation (FID, Inception Score, etc.)

We include a sample_ddp_feature.py script which samples a large number of images from a SiT model in parallel. This script generates a folder of samples as well as a .npz file which can be directly used in evaluator.py which we provided to compute FID, Inception Score and other metrics. For example, to sample 50K images from our pre-trained FlowTurbo model over N GPUs under default ODE sampler settings, run:

torchrun --nnodes=1 --nproc_per_node=N sample_ddp_feature.py \
    --vae_ckpt vae-ema \
    --predictor_ckpt SiT-XL-2-256x256.pt \
    --refiner_ckpt SiT-XL-2-Refiner.pt \
    --num_fid_samples 60 --per_proc_batch_size 20 --cfg_scale 1.5 \
    --tag TEST ;\

Acknowledgments

We would like to express our sincere thanks to the author of SiT for the clear code base.

Citation

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

@article{zhao2024flowturbo,
  title={FlowTurbo: Towards Real-time Flow-Based Image Generation with Velocity Refiner},
  author={Zhao, Wenliang and Shi, Minglei and Yu, Xumin and Zhou, Jie and Lu, Jiwen},
  journal={NeurIPS},
  year={2024}
}

License

This project is under the MIT license. See LICENSE for details.