Discuss matters related to our favourite AI Art generation technology
This is straight from the SDXL beta on discord... can't wait to have this to try on my own PC with Auto1111, I'm sure tweaking the step and cfg settings would fix the texture details. Already loving how much better it is at text and at composition, and from what I've read should be really easy to train and tweak.
Sdxl can do text? Like on the first try?
Exactly, it's definitely better than previous SD versions, but not at Deep Floyd level. It can do short words quite well. Some of the beta versions they were trying were better at text, and some were writing too much text (parts of the prompt would become words in the image). Who knows what we'll end up with, but will be better than SD. This was the best of eight images I made, some had no text, some had ugly lemmings, some had the text with misspellings. I'm sure once we have this in auto or comfy it will be a game changer.
Do you have the prompt for this?
Yes, it was:
Photo of a lemming with a welcome sign, "welcome" written with marker.
No negatives, no fancy adjectives...
I made 8 images and this one was the best (some had no text, some had uglier lemmings, some had misspelled signs) but most were acceptable just not what I wanted. Also the cfg and steps and sampler are random so once we can control that it will be much easier.
Any word on the real world hardware requirements? I'm currently shopping for a machine
Not 100% sure, but from what I read it's not too far from what 2.1 needs. They have even fine tuned it on normal GPUs. I'm not sure if they will have different versions like they had with 2.1 where they released a 512 and a 768 trained version, which would require less VRAM.
I'm hoping for good things!