This dataset, LAION-400M, contains 413M image-text pairs and has subsequently been used "in many papers and experiments." The new dataset, LAION-5B, was collected using a three-stage pipeline.