Research map / Vision-language and multimodal / Multimodal LLMs
ShareGPT4V: Improving Large Multi-Modal Models with Better Captions
Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, Dahua Lin
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 |
|---|---|---|---|
| replace SFT detailed captions with GPT4-Vision captions (3.5% of SFT data) (ablation) | LLaVA-1.5, Qwen-VL-Chat, LLaVA · MME and MMBench | MME / MMBench: MME +222.8/22.0/22.3; MMBench +2.7/1.3/1.5 | Improved |
| add ShareGPT4V captions in SFT only (ablation) | LLaVA-1.5-7B · MME Perception / MMBench / SEED | MME Perception: 1510.7 -> 1542.1; MMBench 64.3 -> 66.8; SEED 66.2 -> 66.7 | Improved |
| add ShareGPT4V-PT captions in pre-training only (ablation) | LLaVA-1.5-7B · MME Perception / MMBench / SEED | MME Perception: 1510.7 -> 1557.2; MMBench 64.3 -> 67.4; SEED 66.2 -> 68.5 | Improved |
| use ShareGPT4V in both pre-training and SFT (ablation) | LLaVA-1.5-7B · MME Perception / MMBench / SEED | MME Perception: 1510.7 -> 1567.4; MMBench 64.3 -> 68.8; SEED 66.2 -> 69.7 | Improved |
| replace BLIP captions with ShareGPT4V-PT captions for the same 558K images (ablation) | LLaVA-1.5-7B pre-training · MME Perception / MMBench / SEED | MME Perception: 1521.6 -> 1539.8; MMBench 66.2 -> 68.3; SEED 66.9 -> 68.9 | Improved |
| joint training with BLIP-558K captions (ablation) | LLaVA-1.5-7B pre-training · MME Perception / SEED | MME Perception: 1516.9 -> 1521.6 | No clear effect |
| unfreeze latter half of vision encoder blocks during pre-training instead of freezing (ablation) | ShareGPT4V-7B · MME Perception / MMBench / SEED | MME Perception: 1515.2 -> 1567.4 | Improved |
| unfreeze more vision blocks than latter half (blocks 6 or 0) (ablation) | ShareGPT4V-7B · MME Perception | MME Perception: 1567.4 -> 1529.5 (block 6); 1545.7 (block 0) | Hurt |
| scale pre-training captions beyond 1000K (ablation) | ShareGPT4V-7B pre-training · MMBench and SEED | accuracy | No clear effect |
| Share-Captioner instead of GPT4-Vision for caption scaling (ablation) | ShareGPT4V-PT generation · human preference over 100 samples | preference percentage: GPT4-Vision 38.2% vs Share-Captioner 35.3% | No clear effect |
Later papers that compared with this method
2 later accepted papers report ShareGPT4V-7B as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Meteor: Mamba-based Traversal of Rationale for Large Language and Vision Models (NeurIPS 2024 · better on 1 of 1 shared result)
- MoAI: Mixture of All Intelligence for Large Language and Vision Models (ECCV 2024 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.88)
- WikiDO: A New Benchmark Evaluating Cross-Modal Retrieval for Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 2026 · similarity 0.88)
- Are We on the Right Way for Evaluating Large Vision-Language Models? (NeurIPS 2024 · similarity 0.88)
- ShareGPT4Video: Improving Video Understanding and Generation with Better Captions (NeurIPS 2024 · similarity 0.88)
- CapsFusion: Rethinking Image-Text Data at Scale (CVPR 2024 · similarity 0.88)
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.
