Research map / Vision-language and multimodal / Human and embodied modeling
Removing Distributional Discrepancies in Captions Improves Image-Text Alignment
Mu Cai, Haotian Liu, Yuheng Li, Yijun Li, Eli Shechtman, Zhe Lin, Yong Jae Lee, Krishna Kumar Singh
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 text-only-classifier data filtering (remove top 30%) (ablation) | LLaVA-score · Winoground | group score: 46.75 -> 47.25 | Improved |
| add data filtering (ablation) | LLaVA-score · MagicBrush | group score: 82.99 -> 87.28 | Improved |
| add data filtering (ablation) | LLaVA-score · SeeTRUE COCO-T2I | ROC-AUC: 81.1 -> 84.9 | Improved |
| filter biased data instead of random subsampling to same size (ablation) | LLaVA-score · MagicBrush | group score: 83.12 -> 87.28 | Improved |
| combine replace and swap negatives (ablation) | LLaVA-score · SugarCrepe add | accuracy: 95.3 -> 97.5 (vs only replace); 91.3 -> 97.5 (vs only swap) | Improved |
| train on swap-only negatives (ablation) | LLaVA-1.5 · MagicBrush | group score: 82.61 -> 76.63 | Hurt |
| fine-tune LLaVA-1.5 on curated data | LLaVA-1.5 · Winoground | group score: 34.25 -> 47.25 | Improved |
| fine-tune LLaVA-1.5 on curated data | LLaVA-1.5 · SeeTRUE EditBench | ROC-AUC: 78.3 -> 77.7 | Mixed |
| fine-tune BLIP2 ITM head on curated data | BLIP2-ITM · SeeTRUE DrawBench | ROC-AUC: 60.8 -> 87.5 | Improved |
| use contrastive loss instead of cross-entropy for fine-tuning (ablation) | LLaVA-score · MagicBrush | group score: 87.28 -> 86.98 | No clear effect |
| increase filtered percentage k beyond 30-40% (ablation) | LLaVA-score · four test datasets | performance | Hurt |
Closest papers by meaning
- Text as Partial Constraint: Core–Residual Alignment for Robust Vision–Language Learning (NeurIPS 2026 · similarity 0.91)
- Mining Fine-Grained Image-Text Alignment for Zero-Shot Captioning via Text-Only Training (AAAI 2024 · similarity 0.90)
- Expanding the Scope of Negatives: Boosting Image-Text Matching with Negatives Distribution Guided Learning (AAAI 2025 · similarity 0.90)
- Extract Free Dense Misalignment from CLIP (AAAI 2025 · similarity 0.89)
- Aligning Text to Image in Diffusion Models is Easier Than You Think (NeurIPS 2025 · similarity 0.89)
- Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language Models (ICML 2024 · 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.
