Research map / Image generation / Text-to-image diffusion
Toward Early Quality Assessment of Text-to-Image Diffusion Models
Huanlei Guo, Hongxin Wei, Bingyi Jing
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 |
|---|---|---|---|
| select top-1 of 5 seeds at t=0.2 using probe-predicted ImageReward | SD2 · MS-COCO val | ImageReward: 0.49 -> 1.59 | Improved |
| select top-1 of 5 seeds at t=0.2 using probe-predicted ImageReward | SD3.5-Medium · MS-COCO val | ImageReward: 1.12 -> 1.83 | Improved |
| select top-1 of 5 seeds at t=0.2 using probe-predicted ImageReward | FLUX.1-dev · MS-COCO val | HPSv2.1: 29.14 -> 31.47 | Improved |
| select with ImageReward probe instead of BLIP-ITM probe (ablation) | SD2 selection · MS-COCO val | ImageReward / CLIPScore: IR 1.09 -> 1.59; CLIPScore 33.96 -> 33.50 | Mixed |
| add early selection (Probe-Select) | SD3-M FID · MS-COCO | FID: 25.26 -> 25.01 | Improved |
| move probe checkpoint from t=0.1 to t=0.2 (ablation) | Probe-Select (SD2) · Spearman with ImageReward | Spearman correlation: 0.70 -> 0.99 | Improved |
| move probe checkpoint from t=0.2 to later (0.3-0.6) (ablation) | Probe-Select · Spearman correlation, all backbones | Spearman correlation: SD2 HPSv2.1 0.64 -> 0.64 | No clear effect |
| increase candidate pool N from 10 to 100 at fixed K=1 (ablation) | Probe-Select (FLUX.1-dev) · 1000 MS-COCO captions | Top-1 ImageReward: 1.36 -> 1.45 | Improved |
| increase continued seeds K (ablation) | Probe-Select (FLUX.1-dev) · 1000 MS-COCO captions | Top-K ImageReward | Hurt |
| select seeds early vs draw more candidates without selection (ablation) | FLUX.1-dev · 1000 MS-COCO captions | ImageReward: 0.95 (no selection) -> 1.45 (Top-1, N=100) | Improved |
Closest papers by meaning
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.91)
- Towards Understanding the Working Mechanism of Text-to-Image Diffusion Model (NeurIPS 2024 · similarity 0.91)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.91)
- R&B: Region and Boundary Aware Zero-shot Grounded Text-to-image Generation (ICLR 2024 · similarity 0.91)
- Image Regeneration: Evaluating Text-to-Image Model via Generating Identical Image with Multimodal Large Language Models (AAAI 2025 · similarity 0.91)
- Prompt-Free Diffusion: Taking "Text" out of Text-to-Image Diffusion Models (CVPR 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.
