Google Brain just announced Imagen (Twitter), which on skimming appears to be not just as good as DALL-E 2 but convincingly better. The main change appears to be reducing the CLIP reliance in favor of a much larger and more powerful text encoder before doing the image diffusion stuff. They make a point of noting superiority on “compositionality, cardinality, spatial relations, long-form text, rare words, and challenging prompts.” The samples also show text rendering fine inside the images as well.
I take this as strong support (already) for my claims 2-3: the problems with DALL-E 2 were not major or deep ones, do not require any paradigm shift to fix, or even any fix, really, beyond just scaling the components almost as-is. (In Kuhnian terms, the differences between DALL-E 2 and Imagen or Make-A-Scene are so far down in the weeds of normal science/engineering that even people working on image generation will forget many of the details and have to double-check the papers.)
EDIT: Google also has a more traditional autogressive DALL-E-1-style 1024px model, “Parti”, competing with diffusion Imagen; it is slightly better in COCO FID than Imagen. It likewise does well on all those issues, with again no special fancy engineering aimed specifically at those issues, mostly just scaling up to 20b.
Google Brain just announced Imagen (Twitter), which on skimming appears to be not just as good as DALL-E 2 but convincingly better. The main change appears to be reducing the CLIP reliance in favor of a much larger and more powerful text encoder before doing the image diffusion stuff. They make a point of noting superiority on “compositionality, cardinality, spatial relations, long-form text, rare words, and challenging prompts.” The samples also show text rendering fine inside the images as well.
I take this as strong support (already) for my claims 2-3: the problems with DALL-E 2 were not major or deep ones, do not require any paradigm shift to fix, or even any fix, really, beyond just scaling the components almost as-is. (In Kuhnian terms, the differences between DALL-E 2 and Imagen or Make-A-Scene are so far down in the weeds of normal science/engineering that even people working on image generation will forget many of the details and have to double-check the papers.)
EDIT: Google also has a more traditional autogressive DALL-E-1-style 1024px model, “Parti”, competing with diffusion Imagen; it is slightly better in COCO FID than Imagen. It likewise does well on all those issues, with again no special fancy engineering aimed specifically at those issues, mostly just scaling up to 20b.