Research map / Image generation / Text-to-image diffusion
Dual-Seed Evolutionary Algorithm for Noise Optimization in Diffusion Models
Yuzheng Tan, Yuan He, Yao Zhu, Tianlin Huo, Huanqian Yan, Hang Su, Shuxin Zhang, Guangneng Hu
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 dual-seed interpolation and guided inversion on top of best-of-N seed selection | RewardSelect · Drawbench (SDv2.1) | Average Gain: 3.8969 vs 0.8010 higher than RewardSelect | Improved |
| add dual-seed interpolation and guided inversion on top of best-of-N seed selection | RewardSelect · Drawbench (SDXL) | Average Gain: 1.3876, exceeding RewardSelect by 0.5173 | Improved |
| apply SOE instead of Golden noise | Golden noise · Drawbench (SDXL) | Average Gain: exceeding Golden noise by 0.5116 | Improved |
| apply SOE instead of Z-sampling | Z-sampling · PickaPic (SDv2.1) | Average Gain: surpassing RewardSelect by 0.4583 and Z-sampling by 2.5879 | Improved |
| apply SOE to vanilla SDv2.1 | SDv2.1 · HPDv2 | ImageReward: 0.3969 -> 1.3405 | Improved |
| apply SOE to vanilla SDXL | SDXL · HPDv2 | ImageReward: 0.9760 -> 1.5637 | Improved |
| apply SOE to vanilla SDXL | SDXL · Drawbench | ClipScore: 0.3291 -> 0.3539 | Improved |
| increase number of seeds per iteration to 3 (ablation) | SOE on SDv2.1 · ImageReward vs vanilla SDXL | ImageReward: requires only 3 seeds to exceed the average performance of vanilla SDXL | Improved |
| increase number of seeds per iteration to 12 (ablation) | SOE on SDv2.1 · ClipScore vs vanilla SDXL | ClipScore: at least 12 seeds are necessary to outperform SDXL on ClipScore | Improved |
| increase number of optimization iterations (ablation) | SOE · ImageReward | ImageReward: increasing the number of rounds improves ImageReward scores | Improved |
| increase number of optimization iterations (ablation) | SOE · ClipScore | ClipScore: yields limited gains on ClipScore | No clear effect |
| apply Initno attention-based noise optimization | SDv2.1 · HPDv2 | Average Gain: -0.1218 | No clear effect |
Closest papers by meaning
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.90)
- RAISE: Requirement-Adaptive Evolutionary Refinement for Training-Free Text-to-Image Alignment (CVPR 2026 · similarity 0.89)
- AdaDiff: Adaptive Step Selection for Fast Diffusion Models (AAAI 2025 · similarity 0.88)
- Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis (CVPR 2025 · similarity 0.88)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.88)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · 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.
