Research map / Image generation / Text-to-image diffusion
Skews in the Phenomenon Space Hinder Generalization in Text-to-Image Generation
Yingshan Chang, Yasi Zhang, Zhiyuan Fang, Ying Nian Wu, Yonatan Bisk, Feng Gao
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 |
|---|---|---|---|
| make training set visually complete and balanced (ablation) | pixel diffusion on What'sUp · What'sUp subsetB (visual skew) | Final accuracy: 18.75 -> 48.64 | Improved |
| increase visual balance while keeping full visual completeness (ablation) | pixel diffusion on What'sUp · What'sUp subsetB (visual skew) | Final accuracy: 10.17 -> 48.64 | Improved |
| make training set linguistically complete and balanced (ablation) | pixel diffusion on What'sUp · What'sUp subsetB (linguistic skew) | Final accuracy: 57.14 -> 67.29 | Improved |
| visual skew versus equal linguistic skew (ablation) | pixel diffusion on What'sUp · What'sUp subsetB | Final accuracy: visual skew worse: 10.17 vs 62.71 at BLC 73 | Hurt |
| increase training coverage (vanilla) without improving completeness or balance (ablation) | synthetic icon diffusion · synthetic icons | testing accuracy | No clear effect |
| increase number of concepts N (ablation) | synthetic icon diffusion · synthetic icons | testing accuracy | Improved |
| improve linguistic completeness and balance (ablation) | synthetic icon diffusion · synthetic icons, final accuracy | final testing accuracy | No clear effect |
| use T5-small text encoder instead of CLIP-B/16 (ablation) | text-to-image diffusion · synthetic icons | testing accuracy | Improved |
| add image positional embeddings (ablation) | pixel diffusion UNet · synthetic icons, N=30 | testing accuracy | Improved |
| use T5-flan-xxl instead of CLIP for spatial role probing (ablation) | text encoder probe · role probing out-of-distribution | out-of-distribution accuracy: 74.5 -> 99.8 | Improved |
| move from pixel to latent diffusion (ablation) | text-to-image diffusion · synthetic and natural images | generalization trends | No clear effect |
Closest papers by meaning
- DreamRelation: Bridging Customization and Relation Generation (CVPR 2025 · similarity 0.87)
- Evaluating Numerical Reasoning in Text-to-Image Models (NeurIPS 2024 · similarity 0.87)
- Progressive Text-to-Image Diffusion with Soft Latent Direction (AAAI 2024 · similarity 0.86)
- Relation Rectification in Diffusion Model (CVPR 2024 · similarity 0.86)
- LLM4GEN: Leveraging Semantic Representation of LLMs for Text-to-Image Generation (AAAI 2025 · similarity 0.86)
- Generating Images of Rare Concepts Using Pre-trained Diffusion Models (AAAI 2024 · similarity 0.86)
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.
