Awesome

Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models

This is the implementation(unconditional generation) of Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models (ECCV2024).

Xiao Liu* , Xiaoliu Guan* , Yu Wu† , and Jiaxu Miao†

Paper arxiv

Text-conditioned implementation is here

1. Introduction

We investigate the memorization phenomenon in the diffusion generative model through the lens of the loss. We propose a novel training framework, Iterative Ensemble Training with Anti-Gradient Control (IET-AGC) to mitigate memorization in diffusion models.

method

2. Environment Setup

Install conda environment

conda env create -f env.yml
conda activate fedes

We use RTX-4090 machines to train the models.

3. Dataset

We evaluate our method on CIFAR-10, CIFAR-100, AFHQ-DOG for unconditional generation.

CIFAR-10 and CIFAR-100

You can download from here, and choose the python version.

AFHQ-DOG

You can download from here. You can only choose the Dog subset.

4. Training and evaluate

Note: Before you run, you must change the paths which include the path of log, generate, datasets and so on.

Default Training

bash com/default.sh

Our Method

bash com/com_c10_v.sh

main.py and main_fed.py, the training files. The default setting is for CIFAR-10 dataset. We used one RTX-4090 GPU to train for almost one day.
all_in.py the evaluate file, includes three parts: generate images, caculate the metric(MQ and FID), and make grid for visualization.

6. Citing

@inproceedings{liu2025iterative,
  title={Iterative ensemble training with anti-gradient control for mitigating memorization in diffusion models},
  author={Liu, Xiao and Guan, Xiaoliu and Wu, Yu and Miao, Jiaxu},
  booktitle={European Conference on Computer Vision},
  pages={108--123},
  year={2025},
  organization={Springer}
}

7. Acknowledgement

We would like to thank this project generously sharing their code, from which our code is adapted.