Research map / Image generation / Fast diffusion sampling
SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders
Enrico Cassano, Riccardo Renzulli, Marco Nurisso, Mirko Zaffaroni, Alan Perotti, 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 |
|---|---|---|---|
| add supervised concept-assignment and decorrelation losses to the SAE (SAEmnesia vs unsupervised SAeUron) | SAE on SD v1.5 · UnlearnCanvas objects | average of UA, IRA, CRA: 82.29 -> 91.51 | Improved |
| supervised one-to-one concept-latent mapping removes the latent-combination search | SAEmnesia · UnlearnCanvas objects | evaluations needed per concept: 210 -> 7 | Improved |
| supervised training for feature centralization | SAEmnesia · Flowers concept score | max feature score: 0.0166 -> 0.0404 | Improved |
| apply supervised SAE to sequential unlearning of nine objects | SAEmnesia · UnlearnCanvas sequential, 9 objects | UA: 64 -> 92.4 | Improved |
| supervised SAE under white-box UnlearnDiffAtk attack | SAEmnesia · UnlearnCanvas objects | UA after attack: 34.20 -> 57.50 | Improved |
| supervised SAE under black-box Ring-A-Bell attack | SAEmnesia · UnlearnCanvas objects | UA after attack: 79.50 -> 97.00 | Improved |
| supervised SAE on nudity removal | SAEmnesia · I2P nudity, SD v1.4 | NudeNet detections: 18 -> 9 | Improved |
| uniform steering multipliers across objects with supervised SAE | SAEmnesia · UnlearnCanvas objects | UA, IRA, CRA across multipliers | Improved |
| using only the top-scoring latent for k-NN classification (ablation) | SAEmnesia · 20 UnlearnCanvas objects | k-NN accuracy | No clear effect |
| compare with SalUn on object unlearning only | SAEmnesia · UnlearnCanvas objects | average of UA, IRA, CRA: 94.28 (SalUn) vs 91.51 | Hurt |
Later papers that compared with this method
1 later accepted paper reports SAEmnesia as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models (ECCV 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- SAeUron — SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders (ICML 2025) · better on 10 of 10 shared results
- SalUn — SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (ICLR 2024) · better on 0 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
- SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders (ICML 2025 · similarity 0.92)
- Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning (NeurIPS 2026 · similarity 0.92)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.90)
- CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.87)
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.
