Research map / Image generation / Autoregressive and latent generation
GUDA: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning
Naoki Murata, Yuhta Takida, Chieh-Hsin Lai, Toshimitsu Uesaka, Bac Nguyen, Stefano Ermon, Yuki Mitsufuji
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ReTrack forget loss instead of ESD in the unlearning operator (ablation) | GUDA · CIFAR-10 group attribution | Top-1 agreement: 0.619 -> 0.727 | Improved |
| replace CLIP-similarity attribution with unlearning-based counterfactual attribution | GUDA vs CLIPA · CIFAR-10 group attribution | Top-1 agreement: 0.662 -> 0.727 | Improved |
| use unlearning approximation instead of LOGO retraining | GUDA · CIFAR-10 group attribution | total wall-clock time (h:mm): 207:47 -> 2:02 | Improved |
| aggregate instance-level gradient scores (TRAK) to group level | TRAK · CIFAR-10 group attribution | Top-1 agreement: 0.118 (10% chance) | Hurt |
| GUDA instead of DAS in head metrics, with faster query time | GUDA · CIFAR-10 group attribution | Top-1 agreement / total time: 0.716 -> 0.727; 35:24 -> 2:02 | No clear effect |
| use GUDA-C anchor redirection instead of CLIPA similarity | GUDA-C · UnlearnCanvas (16 styles) | Top-1 agreement: 0.338 -> 0.456 | Improved |
| instance-level unlearning of a synthesized image (Wang et al.) for group attribution | Wang et al. (2024) · UnlearnCanvas (16 styles) | Top-1 agreement: 0.047 | Hurt |
| use CLIP-weighted instead of uniform retain-style anchor selection (ablation) | GUDA-C · UnlearnCanvas (16 styles) | Spearman: 0.104 -> 0.239 | Improved |
| use retain-style anchor instead of style-removed condition (ablation) | GUDA-C · UnlearnCanvas (16 styles) | Top-1 agreement: 0.400 -> 0.456 | Improved |
| use GUDA-C instead of CLIPA on NDCG@3 | GUDA-C · UnlearnCanvas (16 styles) | NDCG@3: 0.672 -> 0.734 | Improved |
| use GUDA-C unlearning instead of LOGO retraining | GUDA-C · UnlearnCanvas (16 styles) | total wall-clock time (h:mm): 46:08 -> 8:54 | Improved |
Closest papers by meaning
- UnlearnCanvas: Stylized Image Dataset for Enhanced Machine Unlearning Evaluation in Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Diffusion Attribution Score: Evaluating Training Data Influence in Diffusion Models (ICLR 2025 · similarity 0.86)
- The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.86)
- Data Attribution for Text-to-Image Models by Unlearning Synthesized Images (NeurIPS 2024 · similarity 0.85)
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.85)
- EasyRef: Omni-Generalized Group Image Reference for Diffusion Models via Multimodal LLM (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
