Research map / Image generation / Text-to-image diffusion
Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models
Katarzyna Zaleska, Łukasz Popek, Monika Wysoczańska, Kamil Deja
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 |
|---|---|---|---|
| steer after self-attention instead of cross-attention layers (ablation) | ICM steering · Gender debiasing, SD v1.5 | Fairness Discrepancy: 0.365 -> 0.085 | Improved |
| steer after self-attention instead of cross-attention layers (ablation) | ICM steering · Age debiasing, SD v1.5 | Fairness Discrepancy: 0.612 -> 0.273 | Improved |
| steer after self-attention instead of cross-attention layers (ablation) | ICM steering · Race debiasing, SD v1.5 | Fairness Discrepancy: 0.428 -> 0.298 | Improved |
| finetune self-attention layers instead of cross-attention layers (ablation) | ICM finetuning · Gender debiasing, SD v1.5 | Fairness Discrepancy: 0.535 -> 0.463 | No clear effect |
| finetune self-attention layers instead of cross-attention layers (ablation) | ICM finetuning · Age debiasing, SD v1.5 | Fairness Discrepancy: 0.681 -> 0.770 | Hurt |
| select top probe-ranked layers instead of the 3 worst or all layers (ablation) | ICM activation steering · SD v1.5 gender, age, race (visual) | image quality | Improved |
| train probes on general prompts instead of specific prompts (ablation) | ICM steering · Gender, SD v1.5 | FD / FID: FD 0.087 -> 0.089; FID 122.08 -> 119.45 | No clear effect |
| steer only top 20 layers instead of random 20 layers (ablation) | ICM on SDXL · USA president to Barack Obama | Steering success (%): 51.0% -> 83.0% | Improved |
| steer only top 20 layers instead of all layers (ablation) | ICM on SDXL · USA president to Barack Obama | CLIP-I: 0.779 -> 0.893 | Improved |
| steer only top 20 layers instead of all layers (ablation) | ICM on SDXL · USA president to Barack Obama | Steering success (%): 91.0% -> 83.0% | Hurt |
| use ICM steering instead of finetuning baseline | SD v1.5 · Gender debiasing | FID: 161.47 -> 122.08 | Improved |
| restrict steering to 6 best self-attention layers instead of all layers (ablation) | ICM on SANA · gender steering | background preservation (qualitative) | Improved |
Closest papers by meaning
- Precise Parameter Localization for Textual Generation in Diffusion Models (ICLR 2025 · similarity 0.90)
- Unveiling Concept Attribution in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Exploring Local Memorization in Diffusion Models via Bright Ending Attention (ICLR 2025 · similarity 0.89)
- Easing Concept Bleeding in Diffusion via Entity Localization and Anchoring (ICML 2024 · similarity 0.89)
- On Mechanistic Knowledge Localization in Text-to-Image Generative Models (ICML 2024 · similarity 0.88)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.88)
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.
