Research map / Image generation / Text-to-image diffusion
Attention Sinks in Diffusion Transformers: A Causal Analysis
FANGZHENG WU, Brian Summa
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove dynamically identified top-1 attention sink (single layer 12) (ablation) | SD3 inference · GenEval 553 prompts | CLIP-T: -0.5e-3 [-1.7, +0.7] | No clear effect |
| remove top-1 sinks in layers 6, 12 and 18 simultaneously (ablation) | SD3 inference · GenEval 553 prompts | CLIP-T: +0.9e-3 [-0.1, +2.0] | No clear effect |
| increase masked sinks from top-1 to top-5 (ablation) | SD3 inference, layer 12 · GenEval 553 prompts | HPS-v2: -0.0020 (p = 0.001) | Hurt |
| increase masked sinks from top-1 to top-5 (ablation) | SD3 inference, layer 12 · GenEval 553 prompts | CLIP-T: +0.0011 | No clear effect |
| sink masking versus equal-budget random masking at k=10 and k=50 (ablation) | SD3 inference · HPS-v2 preference | HPS-v2 difference-of-differences: -0.005 (k=10), -0.020 (k=50) | Hurt |
| sink masking budget k in {1,5,10,20,50} (ablation) | SD3 inference · CLIP-T alignment | CLIP-T | No clear effect |
| sink masking versus equal-budget random masking (ablation) | SD3 inference, layer 12 · LPIPS perceptual shift | LPIPS: 0.053 -> 0.347 (k=1); 0.104 -> 0.436 (k=5) | Hurt |
| suppress sinks in SDXL mid-block self- or cross-attention (ablation) | SDXL inference · SDXL N=100 | CLIP-T: +0.0004 (attn1); -0.0003 (attn2) | No clear effect |
| full removal sweep (eta from 1 to 0) on score path and value path (ablation) | SD3 inference · dose-response sweep | CLIP-T | No clear effect |
| intervene only in the early denoising phase (ablation) | SD3 inference · phase-specific interventions | CLIP-T: +0.0006 | No clear effect |
| suppress sinks, measured with non-CLIP BLIP2-VQA (ablation) | SD3 inference · GenEval N=553 | BLIP2-VQA score: +0.0001 [-0.0039, +0.0040] | No clear effect |
| increase sink-suppression intensity from A1 to A3 (ablation) | SD3 inference · GenEval | LPIPS vs baseline: 0.189 -> 0.314 | Hurt |
Closest papers by meaning
- When Attention Sink Emerges in Language Models: An Empirical View (ICLR 2025 · similarity 0.89)
- See What You Are Told: Visual Attention Sink in Large Multimodal Models (ICLR 2025 · similarity 0.88)
- Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.87)
- Diffusion PID: Interpreting Diffusion via Partial Information Decomposition (NeurIPS 2024 · similarity 0.87)
- Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models (NeurIPS 2025 · similarity 0.87)
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.
