Research map / Image generation / Fast diffusion sampling
Improving Sample Quality of Diffusion Models Using Self-Attention Guidance
Susung Hong, Gyuseong Lee, Wooseok Jang, Seungryong Kim
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 |
|---|---|---|---|
| add SAG to unconditional ADM | ADM · ImageNet 256x256, LSUN Cat, LSUN Horse | FID, sFID, IS vs recall | Mixed |
| add SAG on top of classifier guidance (ablation) | ADM + CG · ImageNet 128x128 | FID: 2.97 -> 2.58 | Improved |
| add SAG on top of classifier-free guidance (ablation) | DiT-XL/2 · ImageNet 256x256 | FID: 2.27 -> 2.16 | Improved |
| use self-attention masking instead of global blur (ablation) | SAG · ImageNet 128x128 (ADM) | FID: 5.82 -> 5.47 | Improved |
| use self-attention masking instead of random mask (ablation) | SAG · ImageNet 128x128 (ADM) | FID: 5.68 -> 5.47 | Improved |
| use ADM self-attention instead of DINO attention for mask (ablation) | SAG · ImageNet 128x128 (ADM) | FID: 5.63 -> 5.47 | Improved |
| set guidance scale to 0.1 (vs negative or >=0.4) (ablation) | SAG · ImageNet 128x128 (ADM) | FID, sFID, IS | Mixed |
| choose Gaussian blur sigma 3 instead of 1 or 27 (ablation) | SAG · ImageNet 128x128 (ADM) | FID: 5.98 -> 5.47 | Improved |
| set masking threshold to 1.0 instead of 0.7 or 1.3 (ablation) | SAG · ImageNet 128x128 (ADM) | FID: 5.67 -> 5.47 | Improved |
| extract attention from final output layer (ablation) | SAG · ImageNet 128x128 (ADM) | FID: 5.98 -> 5.47 | Improved |
| add SAG to Stable Diffusion with empty prompt | Stable Diffusion · 500 image pairs, human evaluation | human preference | Improved |
| add SAG guidance step | Stable Diffusion · Run-time | run-time: 108.27s -> 186.60s | Hurt |
Closest papers by meaning
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.91)
- Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models (CVPR 2024 · similarity 0.90)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.90)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.90)
- Dynamic Negative Guidance of Diffusion Models (ICLR 2025 · similarity 0.90)
- Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models (NeurIPS 2024 · 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.
