LAION-5B: Dataset for training image-text models
LAION-5B is a 5.85 billion CLIP-filtered image-text pairs starting from Common Crawl using alt-text. It can be used to train SOTA CLIP models of various scale that match the strong zero-shot and robustness performance of the original models trained on closed curated data. CLIP models trained on LAION-400M show competitive zero-shot accuracy compared to CLIP models trained on OpenAI’s WIT. It can also be used to fine-tune generative models like GLIDE, producing samples of good quality.
In this video, I will talk about the following: What is the LAION-5B dataset? How do models trained with LAION dataset perform?
For more details, please look at https://arxiv.org/pdf/2210.08402.pdf
Schuhmann, Christoph, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes et al. "Laion-5b: An open large-scale dataset for training next generation image-text models." Advances in Neural Information Processing Systems 35 (2022): 25278-25294.