Research map / Image generation / Text-to-image diffusion
Alignment-Guided Score Matching for Text-to-Image Alignment in Diffusion Models
Jaa-Yeon Lee, Yeobin Hong, Taesung Kwon, Jong Chul Ye
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 AGSM soft tokens | SD1.5 · COCO val5K | ImageReward (x100): 17.72 -> 34.50 | Improved |
| add AGSM soft tokens | SD1.5 · COCO val5K | FID: 24.59 -> 25.94 | Hurt |
| add AGSM soft tokens | SDXL · COCO val5K | ImageReward (x100): 75.06 -> 84.22 | Improved |
| add AGSM soft tokens | SD3 · COCO val5K | ImageReward (x100): 94.27 -> 103.3 | Improved |
| add AGSM soft tokens | SD3 · COCO val5K | FID: 31.59 -> 34.08 | Hurt |
| add AGSM soft tokens | SD3 · GenEval | counting accuracy: 0.56 -> 0.64 | Improved |
| use negative samples D- with separate soft tokens (psi+, psi-) vs positive only (ablation) | AGSM · COCO val5K | ImageReward (x100): 94.79 -> 103.3 | Improved |
| use separate positive/negative tokens instead of shared tokens (ablation) | AGSM · COCO val5K | ImageReward (x100): 47.33 -> 103.3 | Improved |
| use separate positive/negative tokens instead of shared tokens (ablation) | AGSM · COCO val5K | FID: 31.20 -> 34.08 | Hurt |
| sample with positive tokens only instead of negative tokens for unconditional (ablation) | AGSM · COCO val5K | ImageReward (x100): 84.53 -> 103.3 | Improved |
| use Plackett-Luce multi-candidate loss instead of Bradley-Terry (ablation) | AGSM · COCO val5K | ImageReward (x100): 29.67 -> 34.50 | Improved |
| use Plackett-Luce loss instead of Bradley-Terry (ablation) | AGSM · COCO val5K | FID: 24.76 -> 25.94 | Hurt |
Closest papers by meaning
- Aligning Text to Image in Diffusion Models is Easier Than You Think (NeurIPS 2025 · similarity 0.94)
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.92)
- DyMO: Training-Free Diffusion Model Alignment with Dynamic Multi-Objective Scheduling (CVPR 2025 · similarity 0.91)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.90)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.90)
- Aligning Diffusion Models by Optimizing Human Utility (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.
