Research map / Image generation / Text-to-image diffusion
SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains a BatchTopK sparse autoencoder on cross-attention activations of Stable Diffusion across denoising timesteps, then ablates a few concept-specific SAE features at inference to unlearn objects, styles, and nudity.
Problem
Diffusion models can generate harmful or copyrighted content; existing fine-tuning-based machine unlearning methods are opaque, often only mask concepts, and are vulnerable to adversarial prompts.
Approach
Train one unsupervised sparse autoencoder on activations from all denoising timesteps of a cross-attention block of Stable Diffusion, score features by concept selectivity, and suppress the top-scoring features during inference without modifying model weights or prompts.
Key results
- SAeUron achieves the best average unlearning score on UnlearnCanvas (94.03% avg), with 99.10% IRA and 99.40% CRA on style unlearning.
- On I2P nudity evaluation, SAeUron yields 18 total nudity detections versus 424 for FMN and 111 for MACE, with FID 14.37.
- Under UnlearnDiffAtk on 143 nudity prompts, SAeUron's Post-ASR is 1.40%, the lowest among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAeUron instead of SalUn for UnlearnCanvas | Stable Diffusion v1.5 · UnlearnCanvas | Avg.: 92.43% -> 94.03% | Helped |
| Use SAeUron instead of AdvUn for nudity removal | Stable Diffusion v1.4 · I2P nudity unlearning | Total detected nudity: 28 -> 18 | Helped |
| Train the SAE on half of the benchmark styles rather than the full style set (ablation) | SAeUron · UnlearnCanvas style unlearning, out-of-distribution styles | Avg.: 99.16% -> 72.25% | Mixed |
| Unlearn 49 styles simultaneously (ablation) | SAeUron · UnlearnCanvas, 49/50 styles | IRA: 96.67% | No clear effect |
| Use SAeUron for broad abstract inappropriate-content categories instead of Receler | Stable Diffusion · I2P inappropriate content | Overall inappropriate content ratio: 26.4% -> 44.6% | Hurt |
| Use the object block up.1.1 rather than leaving it unablated (ablation) | Stable Diffusion · Object generation | Object visual quality | Hurt |
| Use the style block up.1.2 rather than leaving it unablated (ablation) | Stable Diffusion · Style generation | Original style presence | Hurt |
Surprising findings
- On the full I2P benchmark judged by Q16, SAeUron underperforms and matches only FMN, which the authors attribute to SAE features capturing concrete visual parts rather than abstract notions like hate or violence.
Limitations
- Performance drops when unlearned and retained concepts are highly similar (e.g., cats and dogs), since shared features get ablated.
- Underperforms on abstract concepts such as hate, harassment, or violence in the full I2P benchmark.
- Requires white-box access; the blocking mechanism could be removed in open-source deployments.
Experimental setup
Datasets: UnlearnCanvas COCO 2014 train captions I2P Anchor prompts from CA
Benchmarks: UnlearnCanvas I2P UnlearnDiffAtk nudity prompts COCO validation FID/CLIPScore
Metrics: UA IRA CRA FID CLIPScore Attack Success Rate Total nudity detections
Baselines: ESD FMN UCE CA SalUn SEOT SPM EDiff SHS AdvUn Receler MACE
Builds on: sparse autoencoders BatchTopK UnlearnCanvas Stable Diffusion UnlearnDiffAtk
Closest papers by meaning
- Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning (NeurIPS 2026 · similarity 0.91)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.90)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.90)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.89)
- CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Erasing Undesirable Concepts in Diffusion Models with Adversarial Preservation (NeurIPS 2024 · 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.