Research map / Image generation / Text-to-image diffusion
Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models
Enrico Cassano, Riccardo Renzulli, Rayyan Ahmed, Stephan Alaniz, Marco Grangetto
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 |
|---|---|---|---|
| replace detected patches (PER) instead of latent steering | SAeUron · UnlearnCanvas object unlearning | Artifacts Rate (%): 57.0 -> 16.3 | Improved |
| replace detected patches (PER) instead of latent steering | SAEmnesia · UnlearnCanvas object unlearning | Artifacts Rate (%): 49.8 -> 15.5 | Improved |
| replace detected patches (PER) instead of latent steering | G-SAE · UnlearnCanvas object unlearning | Artifacts Rate (%): 43.1 -> 22.5 | Improved |
| replace detected patches (PER) instead of latent steering | SAE on SDXL Turbo · UnlearnCanvas object unlearning | Artifacts Rate (%): 61 -> 8 | Improved |
| replace detected patches (PER) instead of latent steering | SAeUron · UnlearnCanvas object unlearning | Global Average (%): 72.47 -> 84.19 | Improved |
| replace detected patches (PER) instead of latent steering | SAEmnesia · UnlearnCanvas object unlearning | Cross-domain retain accuracy (%): 88.48 -> 97.97 | Improved |
| replace detected patches (PER) instead of latent steering | G-SAE · UnlearnCanvas object unlearning | Unlearning Accuracy (%): 78.14 -> 94.02 | Improved |
| replace detected patches (PER) instead of latent steering | SAeUron · UnlearnCanvas object unlearning | Unlearning Accuracy (%): 87.16 -> 85.37 | Hurt |
| use padding p=0 instead of p=1 in mask dilation (ablation) | SAeUron + PER · UnlearnCanvas object unlearning | Unlearning Accuracy (%): 85.37 -> 49.12 | Hurt |
| use padding p=2 instead of p=1 (ablation) | SAEmnesia + PER · UnlearnCanvas object unlearning | Avg of UA/IRA/CRA (%): 93.45 -> 91.70 | Hurt |
| use random replacement instead of nearest concept-free patch (ablation) | SAEmnesia + PER · UnlearnCanvas object unlearning (p=1) | Avg of UA/IRA/CRA (%): 85.07 -> 90.95 | Improved |
| apply PER on top of SAEmnesia for NSFW removal | SAEmnesia · I2P NSFW | exposed body part detections: 47 -> 27 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SAEmnesia — SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders (ICML 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning (NeurIPS 2026 · similarity 0.92)
- CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models (NeurIPS 2026 · similarity 0.90)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.90)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.89)
- One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering (NeurIPS 2026 · similarity 0.89)
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.
