mlsanityData to 14 Oct 2026

Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning

arXiv 2505.24360 ↗First posted 30 May 2025cs.LGcited by 6

On the map: Diffusion models → Diffusion foundations and sampling

also in 1 more
Abstract (arXiv)

Sparse autoencoders are a promising new approach for decomposing language model activations for interpretation and control. They have been applied successfully to vision transformer image encoders and to small-scale diffusion models. Inference-Time Decomposition of Activations (ITDA) is a recently proposed variant of dictionary learning that takes the dictionary to be a set of data points from the activation distribution and reconstructs them with gradient pursuit. We apply Sparse Autoencoders (SAEs) and ITDA to a large text-to-image diffusion model, Flux 1, and consider the interpretability of embeddings of both by introducing a visual automated interpretation pipeline. We find that SAEs accurately reconstruct residual stream embeddings and beat MLP neurons on interpretability. We are able to use SAE features to steer image generation through activation addition. We find that ITDA has comparable interpretability to SAEs.

Key facts. Cited by 6 papers in the map (data through 14 Oct 2026). Placed in Diffusion foundations and sampling, one of 14 directions in Diffusion models. Builds on 28 of its 48 references that are in the map. First posted on arXiv 30 May 2025.

Builds on

28 of 48 references.

Closest ancestors · 6

Same direction, newest first

  1. 2024textitRevelio: Interpreting and leveraging semantic information in diffusion models23 Nov
  2. 2023Unsupervised Discovery of Semantic Latent Directions in Diffusion Models24 Feb
  3. 2022Scalable Diffusion Models with Transformers19 Dec
  4. 2022Diffusion Models already have a Semantic Latent Space20 Oct
  5. 2020Denoising Diffusion Implicit Models6 Oct
  6. 2020Denoising Diffusion Probabilistic Models19 Jun

Other references · 22

  1. 2015U-Net: Convolutional Networks for Biomedical Image Segmentationcited by 19,713
  2. 2023Adding Conditional Control to Text-to-Image Diffusion Modelscited by 5,503
  3. 2022Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flowcited by 3,700
  4. 2017Arbitrary Style Transfer in Real-time with Adaptive Instance Normalizationcited by 2,010
  5. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  6. 2022Toy Models of Superpositioncited by 971
  7. 2023Steering Language Models With Activation Engineeringcited by 922
  8. 2023Extracting Training Data from Diffusion Modelscited by 732
  9. 2024Scaling and evaluating sparse autoencoderscited by 553
  10. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 300
  11. 2023Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Modelscited by 153
  12. 2024The Remarkable Robustness of LLMs: Stages of Inference?cited by 146
10 more
  1. 2021BERT Busters: Outlier Dimensions that Disrupt Transformerscited by 112
  2. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 95
  3. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 92
  4. 2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 80
  5. 2023Nightshade: Prompt-Specific Poisoning Attacks on Text-to-Image Generative Modelscited by 75
  6. 2024A Multimodal Automated Interpretability Agentcited by 63
  7. 2023A General Protocol to Probe Large Vision Models for 3D Physical Understandingcited by 32
  8. 2023Beyond Surface Statistics: Scene Representations in a Latent Diffusion Modelcited by 30
  9. 2025Automated Interpretability Metrics Do Not Distinguish Trained and Random Transformerscited by 30
  10. 2025Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Modelscited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 6.

6 papers.

  1. 2026DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracingcited by 3
  2. 2026Meltdown: Circuits and Bifurcations in Point-Cloud-Conditioned 3D Diffusion Transformerscited by 1
  3. 2026Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Modelscited by 1
  4. 2026The Latent Color Subspace: Emergent Order in High-Dimensional Chaoscited by 0
  5. 2026Robust and Generalizable Safety Steering for Text-to-Image Diffusion Transformerscited by 0
  6. 2026D-Scope: Decomposing and Steering Diffusion Transformers with Sparse Autoencoderscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Probing the Representational Power of Sparse Autoencoders in Vision Models
  2. 2024One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models
  3. 2026Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models
  4. 2024Analyzing (In)Abilities of SAEs via Formal Languages
  5. 2024Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders
  6. 2025Emergence and Evolution of Interpretable Concepts in Diffusion Models
  7. 2024Interpreting Attention Layer Outputs with Sparse Autoencoders
  8. 2026Expander Sparse Autoencoders: Parameter-Efficient Dictionaries for Mechanistic Interpretability