Research map / Vision-language and multimodal / Cross-modal learning
Fragmented Text Is Insufficient for Image Representation: Fine-Grained Correspondence in Multimodal Dataset Distillation
Jingwei Fang, Yaxin Hou, Bo Han, Xu Zhang, Hui LIU, Junhui Hou, Yuheng Jia
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 Intra-pair Text Fusion (ITF) (ablation) | P2DE base · Flickr-30K (100 and 200 pairs) | average recall: 35.1 -> 36.5 | Improved |
| add Inter-pair Contribution Controller (ICC) (ablation) | P2DE base · Flickr-30K (100 and 200 pairs) | average recall: 35.1 -> 39.1 | Improved |
| add ITF on top of ICC (ablation) | P2DE with ICC · Flickr-30K (100 and 200 pairs) | average recall: 39.1 -> 39.8 | Improved |
| fuse captions with image-grounded ITF instead of direct generation (ablation) | P2DE · Flickr-30K (100 pairs) | IR@1: 10.6 -> 12.6 | Improved |
| fuse captions with ITF instead of direct generation (ablation) | P2DE · Flickr-30K (100 pairs) | TR@10: 56.8 -> 57.3 | Improved |
| replace RepBlend contrastive supervision with P2DE | RepBlend · MS-COCO (100 pairs) | IR@10: 22.3 -> 28.8 | Improved |
| replace RepBlend with P2DE | RepBlend · Flickr-30K (200 pairs) | TR@10: 60.0 -> 66.9 | Improved |
| replace RepBlend with P2DE at 500 pairs | RepBlend · Flickr-30K (500 pairs) | IR@10: 55.9 -> 56.6 | No clear effect |
| replace RepBlend with P2DE at larger budget | RepBlend · MS-COCO (1000 pairs) | average recall: 20.5 -> 23.4 | Improved |
| replace RepBlend with P2DE at larger budget | RepBlend · Flickr-30K (1000 pairs) | average recall: 39.9 -> 43.4 | Improved |
| evaluate on ResNet+BERT instead of NFNet+BERT, P2DE vs RepBlend | RepBlend · Flickr-30K (499 pairs, ResNet+BERT eval) | TR@10: 26.2 -> 31.6 | Improved |
| add P2DE to RepBlend (training time) | RepBlend · MS-COCO (100 pairs) | distillation time (s): 25488 -> 27740 | Hurt |
Closest papers by meaning
- Efficient Multimodal Dataset Distillation via Generative Models (NeurIPS 2025 · similarity 0.90)
- Beyond Modality Collapse: Representation Blending for Multimodal Dataset Distillation (NeurIPS 2025 · similarity 0.89)
- CoMM: A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.88)
- CovMatch: Cross-Covariance Guided Multimodal Dataset Distillation with Trainable Text Encoder (NeurIPS 2025 · similarity 0.88)
- OmniCorpus: A Unified Multimodal Corpus of 10 Billion-Level Images Interleaved with Text (ICLR 2025 · similarity 0.87)
- CtrlSynth: Controllable Image Text Synthesis for Data-Efficient Multimodal Learning (ICML 2025 · similarity 0.87)
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.
