I’m not particularly impressed. It’s still making a lot of errors (both in plausibility of output and in following complex instructions eg), and doesn’t seem like a leap over SOTA from last year like Parti—looks like worse instruction-following, maybe better image quality overall. (Of course, people will still be impressed now the way that they should have been impressed last year, because they refuse to believe something exists in DL until they personally can use it, no matter how many samples the paper or website provides to look at.) And it’s still heavily locked-down like DALL-E 2. The prompt-engineering is nice, but people have been doing that for a long time already. The lack of any paper or analysis suggests not much novelty. I’m also not enthused that OA is still screwing with prompts for SJW editorializing.
I’m not particularly impressed. It’s still making a lot of errors (both in plausibility of output and in following complex instructions eg), and doesn’t seem like a leap over SOTA from last year like Parti—looks like worse instruction-following, maybe better image quality overall. (Of course, people will still be impressed now the way that they should have been impressed last year, because they refuse to believe something exists in DL until they personally can use it, no matter how many samples the paper or website provides to look at.) And it’s still heavily locked-down like DALL-E 2. The prompt-engineering is nice, but people have been doing that for a long time already. The lack of any paper or analysis suggests not much novelty. I’m also not enthused that OA is still screwing with prompts for SJW editorializing.