What If We Recaption Billions of Web Images with LLaMA-3?

TL;DR

This paper demonstrates that recaptioning a large-scale web image dataset with an open-source LLaMA-3 model significantly improves the performance of vision-language models in tasks like image retrieval and text-to-image generation.

Contribution

It introduces a scalable recaptioning pipeline using LLaMA-3 to enhance a billion-image dataset, leading to improved model performance across multiple vision-language tasks.

Findings

Enhanced zero-shot cross-modal retrieval performance.

Improved alignment of generated images with complex text queries.

Significant benefits in training vision-language models with Recap-DataComp-1B.

Abstract

Web-crawled image-text pairs are inherently noisy. Prior studies demonstrate that semantically aligning and enriching textual descriptions of these pairs can significantly enhance model training across various vision-language tasks, particularly text-to-image generation. However, large-scale investigations in this area remain predominantly closed-source. Our paper aims to bridge this community effort, leveraging the powerful and \textit{open-sourced} LLaMA-3, a GPT-4 level LLM. Our recaptioning pipeline is simple: first, we fine-tune a LLaMA-3-8B powered LLaVA-1.5 and then employ it to recaption 1.3 billion images from the DataComp-1B dataset. Our empirical results confirm that this enhanced dataset, Recap-DataComp-1B, offers substantial benefits in training advanced vision-language models. For discriminative models like CLIP, we observe enhanced zero-shot performance in cross-modal retrieval tasks. For generative models like text-to-image Diffusion Transformers, the generated images exhibit a significant improvement in alignment with users' text instructions, especially in following complex queries. Our project page is https://www.haqtu.me/Recap-Datacomp-1B/