This model will give you an anime digital painting look with bright, saturated colors, good contrast, and pretty backgrounds. It's an SDXL finetune, so it's good at following prompts and tends to do pretty well at hands. It's decent at nudity, but I didn't train it on any porn so I doubt it can do that.
Description
This is the same train as on v1.0, but trained on top of DreamShaper XL1.0 alpha 2 (shoutouts to @Lykon). It's not definitively better or worse, but it's definitely a bit different, so I'm posting both.
FAQ
Comments (20)
honestly one of the best SDXL models so far, I hope you can continue to improve it. Could you tell me what method you use to create the checkpoints please? Is it a special program? Could you provide me with a link to a tutorial so I can start trying to create my own checkpoints? Thank you so much!
PS: Do you think you can create a refiner at some point to get better results?
Thanks! I'll keep working on it. It was actually trained purely on existing AI art (cherry picked, unlike my initial LoRA test run), so I think my first step is to go in and shore up the quality of the training data a bit.
The program I'm using to finetune is the SDXL development branch of kohya_ss:
https://github.com/bmaltais/kohya_ss/tree/sdxl-dev
I had to make some modifications and run some things from the command line to get it to work, so if you want to go that route, I'd recommend finetuning an SD1.5 checkpoint first just to familiarize yourself with the process. You also need a video card with 24 gigs of ram, which you can rent from a number of sites if you don't own one.
As for a refiner, I'll give that a try, but I'm not 100% sure there's a way to do it at the moment. We're very early after the SDXL release, so I'd imagine these things will get a lot easier in the future.
@_Envy_ Your answer is excellent! Thank you very much for all the information, you really help me a lot to continue learning.
Excellent your work as always!
@mv_ia Sure thing!
If you decide to learn how to finetune networks (or train loras or whatever), feel free to ask me specific questions and I'll help where I can. Unfortunately it's more art than science at this point, so whatever you do will probably end up involving some experimentation. I usually end up trying everything a couple of different ways and going with the best result.
Pruned Version?
pruned != 2GB
This IS the pruned version
SDXL checkpoints are much larger than SD1.5, 2.x
Yeah, non-pruned version is like 13GB. So be glad that it is practically 2 times lower.
@munchkin I think I saw someone on Reddit cut it down further, but there was a noticeable quality drop.
ok, thanks for the explanation :)
Will SD 1.5 Loras and Lycoris and Textual Inversions work with SDXL?
Unfortunately not.
I haven't tried it, but it'll probably just error out. If it doesn't, you'll get a garbled mess.
I have experienced that some textual inversions work with SDXL, ie BadDream and UnrealisticDream work with dreamshaperXL, but only in certain circumstances (square canvas).
@_Envy_ Okay, I'm assuming you'll start rolling out loras for XL now though?
@argh My textural inversions straight up dissapear when i have XL model selected :(
I'm really surprised they work at all, honestly.
@_Envy_ It doesn't. Thanks again for your amazing work, you opened a new paradygm in style by using lora/locon instead of systematic checkpoints.
@sidv Yes, I'm going to be rolling out loras for XL now. :)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.