NerdyRodent

Bonus: AuraFlow 0.2 testing :)

Added 2024-07-28 21:13:55 +0000 UTC

As with the previous AuraFlow 0.1, the latest release is very good at following prompts! This time around the image quality is also better, and there are some limited styles in there too. The main thing seems to be fairly decent text, so long as it's short.

https://huggingface.co/fal/AuraFlow-v0.2/blob/main/aura_flow_0.2.safetensors

The basic workflow is just Aura, including the option upscale too.

The refiner workflow uses an SDXL refiner and finishes with a detailed upscale just because.

Enjoy!

Update:

Added PixArt-Aura-SDXL-US_v5

This one is another experiment because I can never decide which model to use. It's a bit of a beast, and to avoid using any swap space you'll need 38+GB RAM and a 24+GB GPU. It will run with less, but will be slower.

PixArt is great at creating loads of random stuff. From a single fairly short prompt you can get a very wide variety of results. It's better at some subjects and styles than AuraFlow, (it's great at making Zombies and other horror related things, for example) but it can't do text, higher resolutions or really high-quality outputs.

AuraFlow 0.2 is amazing at following prompts and generating text, has pretty decent quality, but doesn't seem as varied as PixArt, and has the subject / style limitations.

SDXL can be creative, has good quality, but can't really do text or follow the prompt well. It also has all the bells and whistle, such as controlnets etc, which can be added for funsies.

Hence, I simply put them all in one workflow :)

It's basically a 3-stage refiner with double upscale. By setting the steps at each stage you can, to some degree, get the best of each world. Want something with text or to really follow the prompt? Use more AuraFlow steps. Want more PixArt? No problem.

Being the slowest, I was also trying to reduce the number of AuraFlow steps to speed things up a bit, and I think you can get some pretty decent ones from Aura with just 16 steps now.

Anyway, it seems fun ;)

Get the new upscaler model from https://huggingface.co/fal/AuraFlow-v0.2/blob/main/aura_flow_0.2.safetensors

Workflows attached below: