Research map / Image generation / Text-to-image diffusion
Diverse Text-to-Image Generation via Contrastive Noise Optimization
Byungjun Kim, Soobin Um, 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 |
|---|---|---|---|
| Optimize initial noise with contrastive loss (CNO) instead of plain DDIM | SD1.5 · MS-COCO val prompts | Vendi Score: 4.6949 -> 4.7855 | Improved |
| Optimize initial noise with CNO | SDXL · MS-COCO val prompts | MSS: 0.2169 -> 0.1623 | Improved |
| Optimize initial noise with CNO | SD1.5 · MS-COCO val prompts | PickScore / Image-Reward: PickScore 21.5081 -> 21.3779; IR 0.2222 -> 0.1284 | Hurt |
| Use gamma=0.01 (strong attraction) (ablation) | CNO · MS-COCO val prompts | Vendi Score: 4.7945 -> 4.7412 | Hurt |
| Use gamma=100.0 (weak attraction) instead of 1.0 (ablation) | CNO · MS-COCO val prompts | Image-Reward: 0.0360 -> 0.0185 | Hurt |
| Lower gamma from 1.0 toward 0.7 (ablation) | CNO · SD1.5 metric variance | sample variance of metrics: Vendi variance 0.076 -> 0.050 (x1e-4) | Improved |
| Pool Tweedie latents to w=4 (ablation) | CNO · PickScore-Vendi Pareto frontier | Pareto frontier | Hurt |
| Use moderate pooling w=8 or 16 instead of full resolution w=64 (ablation) | CNO · PickScore-Vendi Pareto frontier | Pareto frontier | No clear effect |
| Keep full spatial resolution (w=64) instead of w=1 (ablation) | CNO · T2I-CompBench | Image-Reward: -0.2335 -> -0.1697 | Improved |
| Keep full resolution (w=64) instead of w=1 (ablation) | CNO · GenEval | Vendi Score: 4.8374 -> 4.8376 | No clear effect |
| Scale initial noise std by 1.01-1.05 instead of optimizing (ablation) | DDIM · SD1.5 | Vendi Score: 4.7029 -> 4.7058 | No clear effect |
| Scale initial noise std to 1.1 or more (ablation) | DDIM · SD1.5 | PickScore / Vendi: Vendi 4.7029 -> 4.5561 at 1.15 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CADS — CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling (ICLR 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Minority-Focused Text-to-Image Generation via Prompt Optimization (CVPR 2025 · similarity 0.90)
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.90)
- Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.90)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.89)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.88)
- On the Scalability of Diffusion-based Text-to-Image Generation (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.
