D-Scope: Decomposing and Steering Diffusion Transformers with Sparse Autoencoders
cs.CV, cs.AI
Submitted: 2026-09-30
Updated: 2026-09-30
Project page: https://jiahaozhang-public.github.io/d-scope
Terminology
Sources
- Learning Multi-Level Features with Matryoshka Sparse Autoencoders
- Improving Steering Vectors by Targeting Sparse Autoencoder Features
- SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders
- Language Models Can Explain Visual Features via Steering
- LUCID-SAE: Learning Unified Vision-Language Sparse Codes for Interpretable Concept Discovery
- ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
- Concept Steerers: Leveraging K-Sparse Autoencoders for Test-Time Controllable Generations
- CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networks
- LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfer
- Sparse Autoencoders Trained on the Same Data Learn Different Features
- Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders
- Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning
- VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set
- DINOv3
- One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models
- SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features
- AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
- Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models
- UnlearnCanvas: Stylized Image Dataset for Enhanced Machine Unlearning Evaluation in Diffusion Models
Related papers
- Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories
- AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memories
- Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift
- MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation
- TeleOCR: Navigating Document Parsing Across Digital and Camera-Captured Documents
- A Survey on Efficient Vision-Language-Action Models