Data to 7 Oct 2026

SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders

arXiv 2501.18052 ↗First posted 29 Jan 2025cs.LG cs.AIcited by 75

On the map: Diffusion models → Robustness, privacy and adversarial aspects

also in 1 more

LLM safety and security → Vision-language and text-to-image safety

Abstract (arXiv)

Diffusion models, while powerful, can inadvertently generate harmful or undesirable content, raising significant ethical and safety concerns. Recent machine unlearning approaches offer potential solutions but often lack transparency, making it difficult to understand the changes they introduce to the base model. In this work, we introduce SAeUron, a novel method leveraging features learned by sparse autoencoders (SAEs) to remove unwanted concepts in text-to-image diffusion models. First, we demonstrate that SAEs, trained in an unsupervised manner on activations from multiple denoising timesteps of the diffusion model, capture sparse and interpretable features corresponding to specific concepts. Building on this, we propose a feature selection method that enables precise interventions on model activations to block targeted content while preserving overall performance. Our evaluation shows that SAeUron outperforms existing approaches on the UnlearnCanvas benchmark for concepts and style unlearning, and effectively eliminates nudity when evaluated with I2P. Moreover, we show that with a single SAE, we can remove multiple concepts simultaneously and that in contrast to other methods, SAeUron mitigates the possibility of generating unwanted content under adversarial attack. Code and checkpoints are available at https://github.com/cywinski/SAeUron.

Key facts. Cited by 75 papers in the map (data through 7 Oct 2026). Placed in Robustness, privacy and adversarial aspects, one of 14 directions in Diffusion models. Builds on 47 of its 63 references that are in the map. First posted on arXiv 29 Jan 2025.

Builds on

47 of 63 references.

Closest ancestors · 14

Same direction, newest first

2 more

Other references · 33

2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 29,915
2020Denoising Diffusion Probabilistic Modelscited by 20,851
2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,694
2020Denoising Diffusion Implicit Modelscited by 9,655
2015Deep Unsupervised Learning using Nonequilibrium Thermodynamicscited by 7,005
2024GPT-4o System Cardcited by 4,467
2014Going Deeper with Convolutionscited by 4,231
2022LAION-5B: An open large-scale dataset for training next generation image-text modelscited by 4,131
2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
2024Scaling and evaluating sparse autoencoderscited by 530
2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 412

21 more

2022Diffusion Models already have a Semantic Latent Spacecited by 324
2022What the DAAM: Interpreting Stable Diffusion Using Cross Attentioncited by 233
2013k-Sparse Autoencoderscited by 209
2023Selective Amnesia: A Continual Learning Approach to Forgetting in Deep Generative Modelscited by 175
2024BatchTopK Sparse Autoencoderscited by 126
2023Understanding the Latent Space of Diffusion Models through the Lens of Riemannian Geometrycited by 121
2024Automatically Interpreting Millions of Features in Large Language Modelscited by 106
2024Interpreting and Editing Vision-Language Representations to Mitigate Hallucinationscited by 103
2024Get What You Want, Not What You Don't: Image Content Suppression for Text-to-Image Diffusion Modelscited by 62
2024Applying sparse autoencoders to unlearn knowledge in language modelscited by 58
2023Localizing and Editing Knowledge in Text-to-Image Generative Modelscited by 52
2024Scissorhands: Scrub Data Influence via Connection Sensitivity in Networkscited by 50
2024Interpreting Attention Layer Outputs with Sparse Autoencoderscited by 43
2024The Missing Curve Detectors of InceptionV1: Applying Sparse Autoencoders to InceptionV1 Early Visioncited by 37
2024Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localizationcited by 34
2024One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Modelscited by 30
2024textitRevelio: Interpreting and leveraging semantic information in diffusion modelscited by 29
2024Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelinescited by 28
2024On Mechanistic Knowledge Localization in Text-to-Image Generative Modelscited by 27
2024An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generationcited by 23
2024Isometric Representation Learning for Disentangled Latent Space of Diffusion Modelscited by 22

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 24; 2026: 51.

75 papers.

Sort:

By area:

2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Modelscited by 70
2025Set You Straight: Auto-Steering Denoising Trajectories to Sidestep Unwanted Conceptscited by 61
2025I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoderscited by 43
2025A Survey on Mechanistic Interpretability for Multi-Modal Foundation Modelscited by 34
2025Emergence and Evolution of Interpretable Concepts in Diffusion Modelscited by 20
2025Position: Use Sparse Autoencoders to Discover Unknownscited by 17
2025SAUCE: Selective Concept Unlearning in Vision-Language Models with Sparse Autoencoderscited by 12
2025Erasing Concepts, Steering Generations: A Comprehensive Survey of Concept Suppressioncited by 10
2025CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Featurescited by 10
2025Probing the Representational Power of Sparse Autoencoders in Vision Modelscited by 9
2025CytoSAE: Interpretable Cell Embeddings for Hematologycited by 8
2025SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoderscited by 8

38 more

2025SAEdit: Token-level control for continuous image editing via Sparse AutoEncodercited by 8
2026AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoderscited by 8
2025Temporal Concept Dynamics in Diffusion Models via Prompt-Conditioned Interventionscited by 7
2025Interpreting Large Text-to-Image Diffusion Models with Dictionary Learningcited by 6
2026Understanding Emergent Misalignment via Feature Superposition Geometrycited by 6
2026Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarkscited by 6
2025Follow the Flow: On Information Flow Across Textual Tokens in Text-to-Image Modelscited by 4
2025Generative Model Unlearning: A Survey through Target Events, Unlearning Operators, and Evaluation Protocolscited by 4
2025ACE: Attentional Concept Erasure in Diffusion Modelscited by 3
2026UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearningcited by 3
2026Conditioned Activation Transport for T2I Safety Steeringcited by 3
2026OrthoEraser: Coupled-Neuron Orthogonal Projection for Concept Erasurecited by 3
2026RAZOR: Ratio-Aware Layer Editing for Targeted Unlearning in Vision Transformers and Diffusion Modelscited by 3
2026Models Designed to Forget: Machine Unlearning via Key Deletioncited by 3
2026Closed-Form Concept Erasure via Double Projectionscited by 3
2025Memory Self-Regeneration: Uncovering Hidden Knowledge in Unlearned Modelscited by 2
2026ActErase: A Training-Free Paradigm for Precise Concept Erasure via Activation Redirectioncited by 2
2026Selective Fine-Tuning for Targeted and Robust Concept Unlearningcited by 2
2026ELROND: Exploring and decomposing intrinsic capabilities of diffusion modelscited by 2
2026TADA! Tuning Audio Diffusion Models through Activation Steeringcited by 2
2026Improving Sparse Autoencoder with Dynamic Attentioncited by 2
2026Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speechcited by 2
2026DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracingcited by 2
2025Rethinking Sparse Autoencoders: Select-and-Project for Fairness and Control from Encoder Features Alonecited by 1
2025FaCT: Faithful Concept Traces for Explaining Neural Network Decisionscited by 1
2025Deciphering Personalization: Towards Fine-Grained Explainability in Natural Language for Personalized Image Generation Modelscited by 1
2026Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusioncited by 1
2026SPARE: Self-distillation for PARameter-Efficient Removalcited by 1
2026SoftSAE: Dynamic Top-K Selection for Adaptive Sparse Autoencoderscited by 1
2026Disentangled Sparse Representations for Concept-Separated Diffusion Unlearningcited by 1
2026FlowErase-RL: Rethinking Concept Erasure as Reward Optimization in Flow Matching Modelscited by 1
2026Where Concept Erasure Should Occur: Concept-Layer Alignment in Text-to-Video Diffusion Modelscited by 1
2026Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Modelscited by 1
2026Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherencecited by 1
2026Semantic Steering for Controllable Generation: Tuning-Free Concept Erasure in Multimodal Diffusion Transformerscited by 1
2025Beyond the Black Box: Identifiable Interpretation and Control in Generative Models via Causal Minimalitycited by 0
2025SALVE: Sparse Autoencoder-Latent Vector Editing for Mechanistic Control of Neural Networkscited by 0
2026RAIGen: Rare Attribute Identification in Text-to-Image Generative Modelscited by 0

50 most cited of 75 All 75 (JSON)

Similar, not cited

Close in wording, no citation either way