Research map / Vision-language and multimodal / CLIP and zero-shot
BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 ITG loss (ITC+ITM+ITG) to COCO finetuning versus ITC+ITM (ablation) | BLIP-2 image-text retrieval · COCO 5K retrieval | R@1 image-to-text and text-to-image: +0.9 I2T R@1 (84.5 to 85.4); +1.1 T2I R@1 (67.2 to 68.3) | Helped |
| add vision-language representation learning stage before generative learning (ablation) | BLIP-2 with frozen LLMs (OPT, FlanT5) · zero-shot VQA | VQA accuracy: qualitative: without it, both LLM types give substantially lower zero-shot VQA; OPT suffers catastrophic forgetting | Helped |
| use stronger image encoder (ViT-g instead of ViT-L) | BLIP-2 zero-shot VQAv2 with OPT and FlanT5 · VQAv2 zero-shot | VQA accuracy: OPT2.7B 50.1 (ViT-L) to 53.5 (ViT-g); FlanT5XL 62.6 to 63.1 | Helped |
| use larger frozen LLM within family (OPT6.7B instead of OPT2.7B; FlanT5XXL instead of FlanT5XL) | BLIP-2 zero-shot VQA · VQAv2 zero-shot | VQA accuracy: OPT 53.5 to 54.3; FlanT5XL 63.1 to FlanT5XXL 65.2 | Helped |
| use instruction-tuned FlanT5 instead of unsupervised OPT as frozen LLM | BLIP-2 zero-shot VQA · VQAv2 zero-shot | VQA accuracy: ViT-g FlanT5XL 63.1 vs ViT-g OPT2.7B 53.5 (similar scale comparison) | Helped |
| add in-context VQA examples to the frozen LLM prompt | BLIP-2 zero-shot VQA · VQA | VQA accuracy | No clear effect |
| condition Q-Former on the question during VQA finetuning | BLIP-2 VQA finetuning · VQAv2 / GQA | VQA accuracy | No clear effect |
Closest papers by meaning
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Learning by Correction: Efficient Tuning Task for Zero-Shot Generative Vision-Language Reasoning (CVPR 2024 · similarity 0.88)
- Probabilistic Language-Image Pre-Training (ICLR 2025 · similarity 0.88)
- ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models (ICLR 2025 · similarity 0.88)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.88)
- VladVA: Discriminative Fine-tuning of LVLMs (CVPR 2025 · 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.