Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification (CVPR 2025)

Diff-II is a novel and effective generative data augmentation method.

Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification

Yanghao Wang, Long Chen

Updates

• (02/2025) Diff-II was accepted by CVPR 2025.
• (12/2024) Code is released.

Get Start

Environment

Create a conda environment and activate it with the following command:

conda env create -f environment.yaml
conda activate diff-ii

If this takes too long, conda config --set solver libmamba sets conda to use the libmamba solver and could speed up installation.

Download datasets, learned concepts, and suffixes

This may take some time.

bash scripts/download.sh

Run code

We implemented it on 8 x 3090Ti GPUs. Thus, you can adjust the number of GPUs by modifying the code's hyperparameters. Depending on your needs, modify the scripts/run.sh file and run.

bash scripts/run.sh

Citing Diff-II

If you use Diff-II in your research or wish to refer to the baseline results published here, please use the following BibTeX entry.

@inproceedings{wang2025inversion,
  title={Inversion circle interpolation: Diffusion-based image augmentation for data-scarce classification},
  author={Wang, Yanghao and Chen, Long},
  booktitle={2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={25560--25569},
  year={2025},
  organization={IEEE}
}