Research map / Vision-language and multimodal / CLIP and zero-shot
BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation
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 caption generation and filtering to bootstrap the training dataset (ablation) | BLIP pretrained on 14M images · COCO retrieval finetuning | text retrieval R@1: 78.4 -> 80.6 | Helped |
| add filtering without generated captions (ablation) | BLIP dataset bootstrapping · COCO retrieval finetuning | text retrieval R@1: 78.4 -> 79.1 | Helped |
| add generated captions without filtering (ablation) | BLIP dataset bootstrapping · COCO retrieval finetuning | text retrieval R@1: 78.4 -> 79.7 | Helped |
| use nucleus sampling rather than beam search for synthetic captions (ablation) | BLIP CapFilt, 14M images · COCO retrieval finetuning | text retrieval R@1: 79.6 -> 80.6 | Helped |
| decouple captioner and filter parameters rather than sharing them (ablation) | BLIP CapFilt · COCO retrieval finetuning | text retrieval R@1: 79.8 -> 80.6 | Helped |
| share encoder and decoder parameters except for self-attention layers (ablation) | BLIP pretraining · COCO retrieval finetuning | text retrieval R@1: 78.3 -> 78.4 | Helped |
| share self-attention layers between encoder and decoder (ablation) | BLIP pretraining · COCO retrieval finetuning | text retrieval R@1: 78.4 -> 77.3 | Hurt |
| increase the count of noisy web-text samples by replication, without CapFilt (ablation) | BLIP pretraining · COCO retrieval finetuning | text retrieval R@1: 78.4 -> 78.3 | No clear effect |
| continue training the teacher model on bootstrapped data rather than training a new model (ablation) | BLIP · COCO retrieval finetuning | text retrieval R@1: 80.6 -> 80.6 | No clear effect |
| scale pretraining data from 14M to 129M images with CapFilt (ablation) | BLIP with ViT-B · COCO retrieval finetuning | text retrieval R@1: 80.6 -> 81.9 | Helped |
| use a larger ViT-L captioner and filter for CapFilt (ablation) | BLIP with ViT-B pretrained on 129M images · NoCaps validation overall | CIDEr: 106.3 -> 109.6 | Helped |
| increase vision backbone from ViT-B to ViT-L (ablation) | BLIP pretrained on 129M images · COCO retrieval finetuning | text retrieval R@1: 81.9 -> 82.4 | Helped |
Closest papers by meaning
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.89)
- Probabilistic Language-Image Pre-Training (ICLR 2025 · similarity 0.89)
- Learning by Correction: Efficient Tuning Task for Zero-Shot Generative Vision-Language Reasoning (CVPR 2024 · similarity 0.88)
- QLIP: A Dynamic Quadtree Vision Prior Enhances MLLM Performance Without Retraining (ICLR 2026 · similarity 0.88)
- VladVA: Discriminative Fine-tuning of LVLMs (CVPR 2025 · similarity 0.88)
- Probing Visual Language Priors in VLMs (ICML 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.