Dall-E 3

Link post

It seems like we do not yet have a post about Dall-E 3. It was announced a few days ago and can now be tried via bing image creator (long wait times).

I think it is worth mentioning, because it seems like a big jump.

I spend quite a lot of time playing around with Dalle mini, Dall-E 2, Stable Diffusion in its many flavours and it was generally the case that it was very hard to get specifically what I wanted, to the point where I was disillusioned and few of my projects materialised.

A teddy bear and a panda bear holding hands—it took until SDXL for that to work out at least occasionally (the wedding was long over).

Maritime paintings of awesome sailing ships—the rigging was generally completely messed up (finetuning on Montague Dawson and Co didn’t help).

Comic books aka Prince Valiant fan-fic—impossible to reliably get specific poses or situations involving several characters.

Mammoths in space—don’t ask, but I had to do the space hangar in SD and the mammoths in Dall-E and bash it all together.

The Stable Diffusion eco system build around these shortcoming with control nets and finetuned models and increasingly complicated workflows—but now it seems with Dall-E 3 many more things just work right out of the box.

From my few experiments and the examples I have seen elsewhere this might be the point where image generation just works.