On: Dall-E 2

[andybak]

I'm not sure that's "compositing" except in the most abstract sense? But maybe that's the sense in which you mean it.

I'd argue that at no point is there a representation of a "teddy bear" and "a background" that map closely to their visual representation - that are combined.

(I'm aware I'm being imprecise so give me some leeway here)

[astrange → andybak]

This model's predecessor could do image editing with some help:

https://arxiv.org/pdf/2112.10741.pdf

so it could distinguish individual objects from backgrounds. Other ML models can definitely do that; it's called "panoptic segmentation".

[recuter → astrange]

Thank you! Fascinating, I didn't know about panoptic segmentation - that makes things much more interesting.

It really needs to expose the whole pipeline to become truly useful.