D-Scope: Decomposing and Steering Diffusion Transformers with Sparse Autoencoders
On the map: Diffusion models → Image generation, text-to-image and editing
Abstract (arXiv)
Sparse autoencoders (SAEs) reveal visual structure in diffusion transformers (DiTs), but interpreting a feature does not establish whether it can be used to control generation. We introduce D-Scope (Diffusion Scope), a framework that connects feature interpretation to generation control through shared visual evidence. D-Scope aggregates SigLIP~2 embeddings of highly activating image patches into visual centroids. Matching target text descriptions against these visual centroids in the shared image-text embedding space then enables retrieval of individual features without per-feature text annotations. The underlying patches provide evidence for inspecting each selection, while spatially masked interventions test the corresponding decoder direction at varying strengths under fixed generation conditions. We characterize 150 SAEs across two model families and five layers, and introduce a benchmark of 100 target concepts with ten contexts each spanning under-specified and explicit-conflict conditions. Our empirical results show that high reconstruction fidelity can coexist with low dictionary utilization and limited visual-evidence coverage. Under per-case best-of-sweep strength selection, contrastive retrieval yields larger mean regional SigLIP~2 gains than direct retrieval across the tested steering configurations, without consistently improving outside-region preservation. D-Scope provides an inspectable framework for evaluating sparse DiT features through their visual evidence and the effects of their decoder directions on generation. The demo is available at https://jiahaozhang-public.github.io/d-scope/.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Image generation, text-to-image and editing, one of 14 directions in Diffusion models. Builds on 34 of its 37 references that are in the map. First posted on arXiv 30 Sep 2026.
Builds on
34 of 37 references.
Closest ancestors · 4
Same direction, newest first
- 2025Emergence and Evolution of Interpretable Concepts in Diffusion Models21 Apr
- 2025SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation12 Mar
- 2025ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features6 Feb
- 2023Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models20 Nov
Other references · 30
- 2018The Unreasonable Effectiveness of Deep Features as a Perceptual Metriccited by 11,177
- 2022Scalable Diffusion Models with Transformerscited by 6,054
- 2025DINOv3cited by 1,414
- 2025SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Featurescited by 1,188
- 2025SAM 3: Segment Anything with Conceptscited by 906
- 2024Scaling and evaluating sparse autoencoderscited by 553
- 2024PixArt-Sigma: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generationcited by 329
- 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 300
- 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 218
- 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 160
- 2022CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networkscited by 147
- 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
18 more
- 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretabilitycited by 124
- 2025Sparse Autoencoders Trained on the Same Data Learn Different Featurescited by 93
- 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 92
- 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
- 2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 80
- 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
- 2024One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Modelscited by 33
- 2025Mechanistic understanding and validation of large AI models with SemanticLenscited by 29
- 2025TIDE : Temporal-Aware Sparse Autoencoders for Interpretable Diffusion Transformers in Image Generationcited by 16
- 2025VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Setcited by 13
- 2025CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantificationcited by 9
- 2025Concept Steerers: Leveraging K-Sparse Autoencoders for Test-Time Controllable Generationscited by 7
- 2025Interpreting Large Text-to-Image Diffusion Models with Dictionary Learningcited by 6
- 2026Language Models Can Explain Visual Features via Steeringcited by 2
- 2026On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoderscited by 2
- 2025LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfercited by 1
- 2026LUCID-SAE: Learning Unified Vision-Language Sparse Codes for Interpretable Concept Discoverycited by 1
- 2026Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Modelscited by 1
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2025Diffusion Transformers with Representation Autoencoders
- 2026Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models
- 2024Diffusion Models Need Visual Priors for Image Generation
- 2025Muddit: Liberating Generation Beyond Text-to-Image with a Unified Discrete Diffusion Model
- 2025Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models
- 2024CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up
- 2026Learning Representations to Guide Pixel Diffusion Transformers
- 2026Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders