GLIDE: Photorealistic Image Generation, Editing w/ Text-Guided Diffusion Models

GLIDE is NOT Dall-E. Dall-E is a transformer (basically GPT-3), while GLIDE is a diffusion model. While they share some similarities, the major difference is that transformers generate image sequentially from top to bottom, pixel-by-pixel (technically, token-by-token), so one can condition them only by the text and the top of the image. At the same time, diffusion models predict all pixels at the same time, so one can naturally trade compute for result quality (do more inference iterations) and, beside sampling, do other image manipulation tasks, like text-prompted inpainting.

It's easily an impressive enough leap to get hundreds of upvotes. It's about that fact that it takes two clicks (including waiting for arxiv's absurdly slow PDF download speeds), and then some scrolling to get the results.

As someone else said, the blog post that they'll likely release will hit front page.