Research map / Image generation / Text-to-image diffusion
Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback
Jaskirat Singh, Liang Zheng
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 cross-attention updates on top of prompt weighting (ablation) | Ours (PW) · Human evaluation | alignment accuracy (%): 70.28 -> 74.16 | Improved |
| add cross-attention updates on top of prompt weighting (ablation) | Ours (PW) · Human evaluation | normalized human score (%): 89.53 -> 90.25 | Improved |
| add cross-attention updates on top of prompt weighting (ablation) | Ours (PW) · Human evaluation | inference time per image (s): 10.32 s -> 12.24 s | Hurt |
| iterative prompt-weighting refinement vs Attend-and-Excite | Ours (PW) · Human evaluation | alignment accuracy (%): 65.50 -> 70.28 | Improved |
| full method (PW + CA) vs Attend-and-Excite | Stable Diffusion v1.4 · Human evaluation | alignment accuracy (%): 65.50 -> 74.16 | Improved |
| iterative refinement vs plain Stable Diffusion | Stable Diffusion v1.4 · Human evaluation | alignment accuracy (%): 43.66 -> 74.16 | Improved |
| iterative refinement vs plain Stable Diffusion | Stable Diffusion v1.4 · Human evaluation | inference time per image (s): 3.54 s -> 12.24 s | Hurt |
| full method vs Attend-and-Excite on very hard prompts | Stable Diffusion v1.4 · Very hard prompts | alignment accuracy (%): 49.5 -> 62.9 | Improved |
| increase prompt difficulty | all methods · Prompt difficulty levels | alignment accuracy (%) | Hurt |
| Composable-Diffusion vs Stable Diffusion | Stable Diffusion v1.4 · Human evaluation | alignment accuracy (%): 43.66 -> 37.72 | Hurt |
| Structure-Diffusion vs Stable Diffusion | Stable Diffusion v1.4 · Human evaluation | alignment accuracy (%): 43.66 -> 45.23 | No clear effect |
| HPS human-feedback finetuned Stable Diffusion vs original | Stable Diffusion v1.4 · Prompt alignment | semantic alignment | No clear effect |
Closest papers by meaning
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.91)
- Aligning Text to Image in Diffusion Models is Easier Than You Think (NeurIPS 2025 · similarity 0.90)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.90)
- CoMat: Aligning Text-to-Image Diffusion Model with Image-to-Text Concept Matching (NeurIPS 2024 · similarity 0.90)
- Discriminative Probing and Tuning for Text-to-Image Generation (CVPR 2024 · similarity 0.89)
- CHATS: Combining Human-Aligned Optimization and Test-Time Sampling for Text-to-Image Generation (ICML 2025 · similarity 0.89)
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.
