Points to them still relying on a single small text embedding which is lossy instead of full cross-attention; but the fact that it can follow instructions reaso

Points to them still relying on a single small text embedding which is lossy instead of full cross-attention; but the fact that it can follow instructions reasonably well at least means that they've moved off CLIP - phew! It was frustrating arguing with people about DALL-E 2 limitations which were ultimately due to nothing but CLIP.