Research map / Vision-language and multimodal / CLIP and zero-shot
Bootstrapping Vision-Language Learning with Decoupled Language Pre-training
Yiren Jian, Chongyang Gao, Soroush Vosoughi
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 |
|---|---|---|---|
| add P-Former alignment loss to BLIP-2 (4M pairs) | BLIP-2 OPT2.7B · Zero-shot VQAv2 val | accuracy: 46.8 -> 52.6 | Improved |
| add P-Former alignment loss to BLIP-2 (4M pairs) | BLIP-2 OPT2.7B · Zero-shot OK-VQA | accuracy: 25.9 -> 30.0 | Improved |
| add P-Former alignment loss to BLIP-2 (4M pairs) | BLIP-2 OPT2.7B · Zero-shot GQA test-dev | accuracy: 30.5 -> 34.0 | Improved |
| add P-Former alignment loss to BLIP-2 for fine-tuned captioning | BLIP-2 OPT2.7B · COCO Karpathy test | CIDEr: 140.4 -> 141.8 | Improved |
| add P-Former alignment loss to BLIP-2 for NoCaps zero-shot | BLIP-2 OPT2.7B · NoCaps overall | CIDEr: 111.9 -> 115.1 | Improved |
| add alignment loss to stage-1 retrieval pre-training | BLIP-2 stage 1 · Flickr30K image-to-text | R@1: 94.3 -> 93.7 | No clear effect |
| add alignment loss to stage-1 retrieval pre-training | BLIP-2 stage 1 · MSCOCO image-to-text | R@1: 78.4 -> 78.7 | No clear effect |
| add alignment loss only in stage 1 (w1=10, w2=0) (ablation) | BLIP-2 OPT2.7B · Zero-shot VQAv2 val | accuracy: 46.8 -> 51.4 | Improved |
| add alignment loss only in stage 2 (w1=0, w2=100) (ablation) | BLIP-2 OPT2.7B · Zero-shot VQAv2 val | accuracy: 46.8 -> 50.4 | Improved |
| use alignment loss in both stages rather than stage 1 only (ablation) | BLIP-2 OPT2.7B · Zero-shot VQAv2 val | accuracy: 51.4 -> 52.6 | Improved |
| add P-Former alignment loss with Flan-T5XL as LLM (ablation) | BLIP-2 Flan-T5XL · Zero-shot VQAv2 val | accuracy: 48.3 -> 54.9 | Improved |
| train P-Former on 4M VL sentences instead of 12M LAION sentences (ablation) | BLIP-2 + P-Former · Zero-shot VQAv2 val | accuracy: 52.6 -> 51.7 | Hurt |
Closest papers by meaning
- Language Models as Black-Box Optimizers for Vision-Language Models (CVPR 2024 · similarity 0.91)
- VILA: On Pre-training for Visual Language Models (CVPR 2024 · similarity 0.91)
- Frozen Transformers in Language Models Are Effective Visual Encoder Layers (ICLR 2024 · similarity 0.91)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.91)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Weighted Multi-Prompt Learning with Description-free Large Language Model Distillation (ICLR 2025 · similarity 0.90)
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.
