I've been experimenting recently with SDXL 0.9, after seeing it offered in Nightcafe and on clipdrop.co. My first thoughts were that compared to the generators I've been used to (mostly based on Stable Diffusion 2) it tended to produce rather desaturated images with bony, miserable-looking people. But it does have its points. It seems to do a fairly good job with mud prompts: you can ask it for 'young victorian couple covered in mud' and be confident of getting something recognisably muddy, where previous generators I've tried have drawn them clean or just with mud-coloured clothes. And I've had the occasional good paint or slime image out of it too.

Is it perfect? No; some of my tests still produced hands with the wrong number of fingers, or embracing couples whose arms are fused together. But it definitely seems to have a better understanding of prompts than the other generators I've looked at.

And I see that SDXL 1.0 has now been released, so once the daily limit resets I'll have to see how that performs...

I was wondering what the SDXL base model could generate and when we would see the first images of it here. It seems to be the newest trend. It is interesting it applies mud to clothing more easily.

I was considering trying it locally, but I think I need to upgrade my hardware for it. I'm even considering trying out some remote options.

I am still using 1.5 because I am able to train models and generate images quickly. I like generating 512x768 and upscale by 2.5 to 1280x1920. Upscaling will sometimes lose details and textures though.

I'm interested in seeing what 1.0 can generate.

Shouldn't need drastically better hardware for it. Only issue I've found is that it's not really optimised for Automatic1111 in terms of workflow, with the 2-stage model. ComfyUI is an alternative, which has a more flexible workflow. Takes a minute to grok the docs, but it's less daunting than it initially appears!

https://github.com/comfyanonymous/ComfyUI

Here are some quick outputs from 1.0. Certainly there seems to be a niche for AI when the alternative is finding a stately home that lets you have gungings in the drawing room...

AI seems to think that mountain bikes have astonishingly strong joins and no top tubes...

Bikes are one of those things, like hands, that AI doesn't seem to grasp the structure of. With a lot of my experiments I got the impression that the reason my riders had ended up in the mud was that the bike had disintegrated under them...

It is a shame the AI cannot get its head around the bikes, because it is a wonderful slapstick idea - losing control of the bike and ending up in a muddy pool

It's inspired by one of my more vivid muddy childhood memories - trying to ride my bike through a big puddle, getting stuck halfway, and the realisation that I was only wearing sandals and there was no clean way out