Research map / Vision-language and multimodal / Human and embodied modeling
OpenVision 2: A Family of Generative Pretrained Visual Encoders for Multimodal Learning
Yanqing Liu, Xianhang Li, Letian Zhang, Zirui Wang, Zeyu Zheng, Yuyin Zhou, Cihang Xie
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 |
|---|---|---|---|
| remove text encoder and contrastive loss (caption-only training) | OpenVision L/14 · Open-LLaVA-Next (336px) | TextVQA / OCRBench: TextVQA 68.3 -> 68.9; OCR 547 -> 537 | No clear effect |
| remove text encoder and contrastive loss | OpenVision L/14 · Open-LLaVA-Next (336px) | ChartQA: 68.0 -> 62.3 | Hurt |
| remove text encoder and contrastive loss | OpenVision L/14 · LLaVA-1.5 (336px) | TextVQA / OCRBench: TextVQA 61.2 -> 63.0; OCR 339 -> 357 | Improved |
| remove text encoder and contrastive loss | OpenVision L/14 · Training time (TPU v4-512) | hours: 83 -> 57 | Improved |
| remove text encoder and contrastive loss | OpenVision L/14 · Peak memory, batch 2k | GB per chip: 24.5 -> 13.8 | Improved |
| add CLIPA low-res optimization and token masking (ablation) | CapPa baseline · Training time L/14 224px | hours: 217 -> 55 | Improved |
| replace alt-text with synthetic ReCap captions (ablation) | OpenVision 2 · Caption type ablation | TextVQA / OCRBench: TextVQA 51.8 -> 56.5; OCR 238 -> 303 | Improved |
| use ReCap v2 (alt-text conditioned) instead of v1 captions (ablation) | OpenVision 2 · Caption type ablation | TextVQA / OCRBench: TextVQA 56.9 -> 56.5; OCR 291 -> 303 | Mixed |
| mask visual tokens (keep 35%) instead of keeping 100% (ablation) | OpenVision 2 · Token keep ratio ablation | TextVQA / OCRBench: TextVQA 53.8 -> 56.9; OCR 254 -> 291 | Improved |
| keep only 10% of tokens instead of 35% (ablation) | OpenVision 2 · Token keep ratio ablation | TextVQA / OCRBench: TextVQA 56.9 -> 55.6; OCR 291 -> 276 | Hurt |
| increase max caption length from 16 to 128 tokens (ablation) | OpenVision 2 · Max caption length ablation | TextVQA / OCRBench: TextVQA 54.8 -> 56.5; OCR 269 -> 303 | Improved |
| add auxiliary MAE pixel reconstruction loss (ablation) | OpenVision 2 · multimodal benchmarks | benchmark scores | Hurt |
Closest papers by meaning
- Multimodal Autoregressive Pre-training of Large Vision Encoders (CVPR 2025 · similarity 0.90)
- iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models (ICLR 2026 · similarity 0.90)
- Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs (NeurIPS 2026 · similarity 0.90)
- Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders (ICLR 2026 · similarity 0.90)
- CoVFT: Context-aware Visual Fine-tuning for Multimodal Large Language Models (CVPR 2026 · similarity 0.89)
- Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation (NeurIPS 2026 · 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.
