If you ever actually look into Unstable diffusion Kickstarter, you’ll notice that they’re not actually claiming they’ll manually label a dataset the size of Lai

If you ever actually look into Unstable diffusion Kickstarter, you’ll notice that they’re not actually claiming they’ll manually label a dataset the size of Laion-5B - that’s a much bigger task than what you seem to think it is. Even if a million people are labeling images, without any overlap, 5 billion images would mean each of them has to label 5000 images each. What Unstable diffusion folks seem to be doing is that they’re using a few thousand labeled images to train a caption generation model and then use it to create a huge multimodal dataset with text and high quality images.