GUDA: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning

arXiv | Accepted at ICML 2026

This repository contains the public reproduction code for the GUDA experiments.

GUDA (Group Unlearning-based Data Attribution) approximates Leave-One-Group-Out counterfactual models for diffusion models by unlearning each group from a shared full-data model. This repository covers the CIFAR-10 class attribution and UnlearnCanvas style attribution experiments from the paper.

For end-to-end reproduction, including artifact dependencies and validation checks, start with REPRODUCTION.md.

Paper

• Title: GUDA: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning
• Authors: Naoki Murata¹, Chieh-Hsin Lai¹, Yuhta Takida¹, Toshimitsu Uesaka¹, Bac Nguyen¹, Stefano Ermon², Yuki Mitsufuji^1,3
• Affiliations: ¹ Sony AI, ² Stanford University, ³ Sony Group Corporation
• Contact: naoki.murata@sony.com
• Venue: ICML 2026
• Preprint: https://arxiv.org/abs/2601.22651

Repository Layout

CIFAR10/          CIFAR-10 DDPM training, LOGO, GUDA-U, and ranking evaluation
UnlearnCanvas/    SD 1.5 fine-tuning, LOGO, GUDA-C, and style attribution
env/              Minimal Python dependency lists
docs/examples/    Runnable command examples
REPRODUCTION.md   End-to-end reproduction guide and artifact dependency map

Datasets, checkpoints, generated images, and precomputed caches are not distributed. The code expects users to provide or produce those artifacts under local data/, checkpoints/, outputs/, and results/ directories.

Release Scope

This repository is intended to contain only source code written for the GUDA experiments and prompt metadata used for the UnlearnCanvas evaluation. It does not include third-party source code, model weights, checkpoints, generated images, precomputed caches, or image datasets.

External models and datasets referenced by the commands, including Stable Diffusion 1.5, CIFAR-10, and UnlearnCanvas images, must be obtained separately by users under their respective licenses, terms, and access policies. The distributed UnlearnCanvas files are prompt and descriptor metadata only.

Environment

Use separate environments for the CIFAR-10 and UnlearnCanvas pipelines. Install PyTorch first with the CUDA wheel appropriate for your system, then install the pipeline requirements:

python3 -m venv .venv-cifar10
source .venv-cifar10/bin/activate
python3 -m pip install --upgrade pip
# Install CUDA-enabled torch/torchvision for your system first.
python3 -m pip install -r env/cifar10.requirements.txt

python3 -m venv .venv-uc
source .venv-uc/bin/activate
python3 -m pip install --upgrade pip
# Install CUDA-enabled torch/torchvision for your system first.
python3 -m pip install -r env/unlearncanvas.requirements.txt

GPU reproduction requires a CUDA-enabled PyTorch build, not the CPU-only wheel. See env/README.md for known-working reference environments and more detailed setup notes.

CIFAR-10 Reproduction Order

Run from CIFAR10/ with PYTHONPATH=..

1. Train the full unconditional CIFAR-10 model.
2. Train the LOGO models, one model per held-out class. This is expensive and is expected to be run by the user.
3. Train GUDA-U unlearned models with ReTrack for classes 0-9.
4. Optionally train GUDA-U variants with ESD for the ReTrack-vs-ESD comparison.
5. Generate query images and compute LOGOA / GUDA attribution scores.
6. Compare LOGOA and GUDA rankings with the ranking metrics.
7. Run ReTrack ablations over unlearning epochs, learning rate, forget/preserve weighting, and nearest-neighbor count.

See docs/examples/cifar10_commands.md and the batch examples under CIFAR10/scripts/run_*.sh. See REPRODUCTION.md for the full artifact graph.

UnlearnCanvas Reproduction Order

Run from UnlearnCanvas/ with PYTHONPATH=..

1. Prepare the UnlearnCanvas images locally. Prompt metadata used by the paper is included under UnlearnCanvas/data/.
2. Fine-tune SD 1.5 on the 16 target styles.
3. Train 16 LOGO models, one model per target style. This is expensive and is expected to be run by the user.
4. Train GUDA-C with the AWSS anchor strategy.
5. Generate the 320 query images and compute LOGOA / GUDA attribution scores.
6. Compare LOGOA and GUDA rankings.
7. Run anchor strategy ablations: AWSS, uniform sampling, and style removed.

See docs/examples/unlearncanvas_commands.md and the batch examples under UnlearnCanvas/scripts/run_*.sh. See REPRODUCTION.md for the full artifact graph and the 16 paper-faithful target styles.

Maintenance

This repository is maintained for reproduction support.

Citation

The final ICML citation will be updated when available. For now, please cite the arXiv version:

@misc{murata2026guda,
  title        = {{GUDA}: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning},
  author       = {Murata, Naoki and Takida, Yuhta and Lai, Chieh-Hsin and Uesaka, Toshimitsu and Nguyen, Bac and Ermon, Stefano and Mitsufuji, Yuki},
  year         = {2026},
  eprint       = {2601.22651},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url          = {https://arxiv.org/abs/2601.22651}
}

License

This project is licensed under the MIT License. See LICENSE for details.