BLIP3-KALE: Knowledge Augmented Large-Scale Dense Captions

TL;DR

BLIP3-KALE introduces a massive dataset of 218 million image-text pairs that combines synthetic dense captions with web-scale alt-text, enhancing factual grounding and training more capable vision-language models.

Contribution

The paper presents a novel large-scale dataset, KALE, and a two-stage approach for generating knowledge-augmented captions to improve vision-language model training.

Findings

Improved performance on vision-language tasks using KALE-trained models.

Demonstrated the effectiveness of knowledge-augmented captions for factual grounding.

Released a publicly accessible dataset for further research.

Abstract

We introduce BLIP3-KALE, a dataset of 218 million image-text pairs that bridges the gap between descriptive synthetic captions and factual web-scale alt-text. KALE augments synthetic dense image captions with web-scale alt-text to generate factually grounded image captions. Our two-stage approach leverages large vision-language models and language models to create knowledge-augmented captions, which are then used to train a specialized VLM for scaling up the dataset. We train vision-language models on KALE and demonstrate improvements on vision-language tasks. Our experiments show the utility of KALE for training more capable and knowledgeable multimodal models. We release the KALE dataset at https://huggingface.co/datasets/Salesforce/blip3-kale