Research map / Image generation / Text-to-image diffusion
Emergence and Evolution of Interpretable Concepts in Diffusion Models
Berk Tinaz, Zalan Fabian, Mahdi Soltanolkotabi
Read the paperCheck your idea against 137,217 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 per-block, per-timestep TopK sparse autoencoders on Stable Diffusion v1.4 cross-attention residual updates, builds a vision-only concept dictionary, and uses it to predict early image layout and to steer composition and style at different diffusion stages.
Problem
placeholder
Approach
placeholder
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply spatially targeted concept interventions at t = 1.0 rather than using random quadrant placement | Stable Diffusion v1.4 · Spatially targeted interventions | spatial edit success: 0.25 -> 0.80 | Improved |
| Apply spatially targeted concept interventions at t = 0.5 rather than random quadrant placement | Stable Diffusion v1.4 · Spatially targeted interventions | spatial edit success: 0.25 -> 0.35 | No clear effect |
| Apply spatially targeted concept interventions at t = 0.0 rather than random quadrant placement | Stable Diffusion v1.4 · Spatially targeted interventions | spatial edit success: 0.25 -> 0.23 | No clear effect |
| Apply global style-concept interventions at t = 0.5 rather than at t = 0.0 | Stable Diffusion v1.4 · Global interventions | edit success: 0.78 -> 0.93 | Improved |
| Apply global style-concept interventions at t = 0.5 rather than at t = 1.0 | Stable Diffusion v1.4 · Global interventions | ΔCLIP: +0.007 -> +0.021 | Improved |
| Apply global style-concept interventions at t = 1.0 | Stable Diffusion v1.4 · Image style | style change | Hurt |
| Apply global style-concept interventions at t = 0.5 | Stable Diffusion v1.4 · Image style | style change: 0.199 -> 0.220 CLIP similarity; +0.021 ΔCLIP | Improved |
| Apply global style-concept interventions at t = 0.0 | Stable Diffusion v1.4 · Image style | style change: +0.006 ΔCLIP | No clear effect |
| Progress from t = 1.0 to t = 0.0 during reverse diffusion | Stable Diffusion v1.4 SAE concepts · Concept dictionary | concept cohesion: 0.588 -> 0.664 | Improved |
| Progress from t = 1.0 to t = 0.0 during reverse diffusion | Stable Diffusion v1.4 SAE concepts · Concept dictionary | concept separability: 0.433 -> 0.344 | Improved |
| Use text-conditioned rather than null-text-conditioned SAE features | Stable Diffusion v1.4 composition prediction · Image composition segmentation | segmentation accuracy | Improved |
| Use up_block rather than down_block SAE features | Stable Diffusion v1.4 composition prediction · Image composition segmentation | segmentation accuracy | Improved |
Closest papers by meaning
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.93)
- CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models (NeurIPS 2026 · similarity 0.91)
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.90)
- Diffusion PID: Interpreting Diffusion via Partial Information Decomposition (NeurIPS 2024 · similarity 0.90)
- One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Unveiling Concept Attribution in Diffusion Models (NeurIPS 2025 · similarity 0.90)
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.
