Research map / Image generation / Fast diffusion sampling
Toward the Frontiers of Reliable Diffusion Sampling via Adversarial Sinkhorn Attention Guidance
Kwanyoung 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 |
|---|---|---|---|
| replace heuristic attention perturbation (PAG) with adversarial Sinkhorn attention | SDXL unconditional sampling · MS-COCO unconditional | FID: 108.63 -> 92.01 | Improved |
| replace SEG with ASAG | SDXL unconditional sampling · MS-COCO unconditional | FID: 95.43 -> 92.01 | Improved |
| add ASAG to CFG | SDXL conditional · MS-COCO conditional | FID: 28.15 -> 23.30 | Improved |
| add ASAG to CFG | SD3 conditional · MS-COCO conditional | FID: 24.19 -> 22.87 | Improved |
| add ASAG to CFG | SDXL · DrawBench | ImageReward: 0.196 -> 0.316 | Improved |
| use ASAG instead of SEG on HPSv2 | SDXL · DrawBench | HPSv2: 27.06 -> 27.08 | No clear effect |
| use minimizing-similarity cost QK^T instead of similarity-maximizing 1 - QK^T (ablation) | ASAG · MS-COCO unconditional (SDXL) | FID: 111.53 -> 92.01 | Improved |
| use similarity-maximizing Sinkhorn cost versus vanilla sampling (ablation) | SDXL unconditional · MS-COCO unconditional (SDXL) | FID: 122.07 -> 111.53 | Improved |
| use Sinkhorn plan instead of extreme uniform plan (ablation) | ASAG · MS-COCO unconditional (SDXL) | Inception Score: 9.710 -> 10.54 | Improved |
| use Sinkhorn plan instead of extreme uniform plan (ablation) | ASAG · MS-COCO unconditional (SDXL) | FID: 92.11 -> 92.01 | No clear effect |
| add ASAG guidance instead of PAG/SEG at similar cost | SDXL · inference time per prompt | seconds: 1.198 -> 1.551 | Hurt |
| increase ASAG guidance scale beyond 1.5 (ablation) | ASAG · MS-COCO | generation quality | Hurt |
Closest papers by meaning
- Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Perturbing Attention Gives You More Bang for the Buck: Subtle Imaging Perturbations That Efficiently Fool Customized Diffusion Models (CVPR 2024 · similarity 0.89)
- Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models (NeurIPS 2025 · similarity 0.88)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.87)
- Dynamic Negative Guidance of Diffusion Models (ICLR 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.
