AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

ScaleCap: Scalable Image Captioning via Dual-Modality Debiasing

Long Xing, Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jinsong Li, Shuangrui Ding, Weiming Zhang, Nenghai Yu, Jiaqi Wang, Feng Wu and 1 more

ICLR 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Concept-guided image captioning

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 changedWhereMeasuredEffect
further pretrain on ScaleCap-450k captions instead of DenseFusion-450kLLaVA-NeXT style LVLM · 11-benchmark average, Qwen2.5-7Baverage score: 63.0 -> 64.7Improved
further pretrain with ScaleCap captions on the same DenseFusion images (ablation)Qwen2.5-3B LVLM · 11-benchmark averageaverage score: 58.9 -> 60.2Improved
add both object and position instructions (vs object only) (ablation)ScaleCap · Prism subset avgaverage: 58.8 -> 61.5Improved
add object instructions (vs position only) (ablation)ScaleCap · Prism subset avgaverage: 57.4 -> 61.5Improved
scale summarization LLM from Qwen2-7B to Qwen2-72B (ablation)ScaleCap · MMVetscore: 43.6 -> 58.8Improved
scale LVLM from 7B to 72B (ablation)ScaleCap · Prism subset avgaverage: 58.5 -> 58.7No clear effect
scale LVLM from 2B to 7B (ablation)ScaleCap · Prism subset avgaverage: 50.2 -> 58.5Improved
increase number of heuristic instructions N (ablation)ScaleCap · MMVet (Prism)accuracy: 51.8 -> 58.8Improved
add golden-sentence contrastive filtering (ablation)LLaVA-v1.5 7B · CHAIRCHAIR_S: 48.8 -> 33.6Improved
add golden-sentence contrastive filtering (ablation)Qwen2-VL 7B · CHAIRCHAIR_S: 44.2 -> 25.8Improved
use ScaleCap captions with Qwen2-VL-7B instead of Qwen2-VL-72B generic captionsPrism pipeline · Prism 6-benchmark averageaverage accuracy: 56.0 -> 58.2Improved
ScaleCap pretraining vs DenseFusion on ChartQA/MMMU-like individual benchmarksQwen2.5-7B LVLM · MMStar and MMMUscore: MMStar 49.2 -> 48.8; MMMU 44.3 -> 46.1Mixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy