Distilling Pretrained Diffusion-Based Generative Models with SiD

This repository now contains two distinct branches:

1. Main Branch: Focused on Score identity Distillation (SiD) as described in our ICML 2024 paper.
2. SiDA Branch: Focused on Adversarial Score identity Distillation (SiDA), with additional techniques for distilling EDM2 models, as described in our ICLR 2025 paper.

For more details on the SiDA branch, visit: https://github.com/mingyuanzhou/SiD/tree/sida.

---

About This Repository

This repository contains the code necessary to replicate the findings of our ICML 2024 paper titled "Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation," available at arXiv:2404.04057. The technique, Score identity Distillation (SiD), is used to distill pretrained EDM diffusion models.

Branch Overview

• Main Branch:

  • Includes the implementation of Score identity Distillation (SiD) as described in the ICML 2024 paper.

• SiDA Branch:

  • Provides the latest advancements in Adversarial Score identity Distillation (SiDA), including techniques for distilling EDM2 models and leveraging adversarial learning.

Citations

If you find our work useful or incorporate our findings in your own research, please consider citing our paper:

@inproceedings{zhou2024score,
  title={Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation},
  author={Mingyuan Zhou and Huangjie Zheng and Zhendong Wang and Mingzhang Yin and Hai Huang},
  booktitle={International Conference on Machine Learning},
  year={2024}
}

We also have a follow-up paper, available at https://arxiv.org/abs/2406.01561 since June 2024 and published at ICLR 2025, that extends our SiD methodology to distill Stable Diffusion models for one-step text-to-image generation:

@article{zhou2024long,
title={Long and Short Guidance in Score identity Distillation for One-Step Text-to-Image Generation},
author={Mingyuan Zhou and Zhendong Wang and Huangjie Zheng and Hai Huang},
journal={ArXiv 2406.01561},
url={https://arxiv.org/abs/2406.01561},
year={2024}
}

State-of-the-art Performance

SiD operates as a data-free distillation method but still demonstrates superior performance compared to the teacher EDM model across most datasets, with the notable exception of ImageNet 64x64. It outperforms all previous diffusion distillation approaches—whether one-step or few-step, data-free or training data-dependent—in terms of generation quality. This achievement sets new standards for efficiency and effectiveness in diffusion distillation.

It achieves the following Fréchet Inception Distances (FID):

Prerequisites

Before you begin, ensure you have met the following requirements:

• You have installed the latest version of Conda.
• You have a Windows/Linux/Mac machine.

Installation

To install the necessary packages and set up the environment, follow these steps:

Clone the Repository

First, clone the repository to your local machine:

git clone https://github.com/mingyuanzhou/SiD.git
cd SiD

Create the Conda Environment

To create the Conda environment with all the required dependencies, run:

conda env create -f environment.yaml

This command will read the environment.yaml file in the repository, which contains all the necessary package information.

Activate the Environment

After creating the environment, you can activate it by running:

conda activate sid

Prepare the Datasets

Follow the instructions detailed in the EDM codebase to prepare the training datasets. Once prepared, place them into the /data/datasets/ folder:

• cifar10-32x32.zip
• imagenet-64x64.zip
• ffhq-64x64.zip
• afhqv2-64x64.zip

Note: Although a training dataset is not necessary for distilling the pretrained EDM model, it is used in our code to calculate evaluation metrics such as FID and Inception Score. Optionally, you can create a dummy dataset and either disable the evaluation code if you wish to run the SID distillation code without these metrics, or provide an npz file of the training dataset if you need to compute these metrics.

Usage

Training

After activating the environment, you can run the scripts or use the modules provided in the repository. Example:

sh run_sid.sh 'cifar10-uncond'

Adjust the --batch-gpu parameter according to your GPU memory limitations. The default setting for cifar10-uncond consumes less than 10 GB of memory per GPU.

Generation

Generate example images:

Generate images and save them as out/*.png and out.npz

• Using a single GPU

python generate_onestep.py --outdir=image_experiment/sid-train-runs/out --seeds=0-63 --batch=64 --network=<network_path>

• Using multiple GPU

torchrun --standalone --nproc_per_node=2 generate_onestep.py --outdir=image_experiment/sid-train-runs/out --seeds=0-999 --batch=64 --network=<network_path>

Evaluations

For ImageNet, there are two different versions of the training data, each associated with its own set of reference statistics. To ensure apples-to-apples comparisons between EDM and its distilled generators with other diffusion models, imagenet-64x64.npz should be used for computing FID (Fréchet Inception Distance). Conversely, for computing Precision and Recall, VIRTUAL_imagenet64_labeled.npz should be utilized.

imagenet-64x64.npz is available at NVIDIA.

VIRTUAL_imagenet64_labeled.npz is available at OpenAI.

Use sid_generator.py to generate and save 50,000 images, and compute FID using the saved images

Use a single GPU

python sid_generate.py --outdir=image_experiment/out --seeds=0-49999 --batch=128 --network='https://huggingface.co/UT-Austin-PML/SiD/resolve/main/cifar10-uncond/alpha1.2/network-snapshot-1.200000-403968.pkl' --ref=https://nvlabs-fi-cdn.nvidia.com/edm/fid-refs/cifar10-32x32.npz 

Use four GPUs

torchrun --standalone --nproc_per_node=4 sid_generate.py --outdir=out --seeds=0-49999 --batch=128 --network='https://huggingface.co/UT-Austin-PML/SiD/resolve/main/imagenet64/alpha1.2/network-snapshot-1.200000-939176.pkl' --ref=https://nvlabs-fi-cdn.nvidia.com/edm/fid-refs/imagenet-64x64.npz

Use sid_metrics.py to perform 10 random trials, each trial computes the metrics using 50,000 randomly generated images

Compute FID and/or IS

torchrun --standalone --nproc_per_node=4  sid_metrics.py  --cond=False --metrics='fid50k_full,is50k' --network='https://huggingface.co/UT-Austin-PML/SiD/resolve/main/cifar10-uncond/alpha1.2/network-snapshot-1.200000-403968.pkl' --data='/data/datasets/cifar10-32x32.zip'

torchrun --standalone --nproc_per_node=4  sid_metrics.py  --cond=True --metrics='fid50k_full' --network='https://huggingface.co/UT-Austin-PML/SiD/resolve/main/imagenet64/alpha1.2/network-snapshot-1.200000-939176.pkl' --data='/data/datasets/imagenet-64x64.zip' --data_stat='https://nvlabs-fi-cdn.nvidia.com/edm/fid-refs/imagenet-64x64.npz' 

Compute Precision and Recall for ImageNet

torchrun --standalone --nproc_per_node=4  sid_metrics.py  --cond=True --metrics='pr50k3_full' --network='https://huggingface.co/UT-Austin-PML/SiD/resolve/main/imagenet64/alpha1.2/network-snapshot-1.200000-939176.pkl' --data='/data/datasets/imagenet-64x64.zip' --data_stat='https://openaipublic.blob.core.windows.net/diffusion/jul-2021/ref_batches/imagenet/64/VIRTUAL_imagenet64_labeled.npz'

Checkpoints of one-step generators produced by SiD

The one-step generators produced by SiD are provided in huggingface/UT-Austin-PML/SiD

Acknowledgements

We extend our gratitude to the authors of the EDM paper for sharing their code, which served as the foundational framework for developing SiD. The repository can be found here: NVlabs/edm.

Additionally, we are thankful to the authors of the Diff Instruct paper for making their code available. Their contributions have been instrumental in integrating the evaluation pipeline into our training iterations. Their repository is accessible here: pkulwj1994/diff_instruct.

Code Contributions

• Mingyuan Zhou: Led the project and wrote the majority of the code.
• Huangjie Zheng, Zhendong Wang, Hai Huang: Worked closely with Mingyuan Zhou, co-developing essential components and writing various subfunctions.

Contact

If you want to contact me you can reach me at mingyuanzhou@gmail.com.

License

This project uses the following license: Apache-2.0 license.