Pre-release

I said to myself, fuck it, no captions, no trigger word and well... It works, kind of... I've never used a refiner before, i'd assume it works something like this. I learned from making this lora how much words impact the result of the image. Also now I know concepts like this only need 1200 steps, only characters need 5000+ steps. I let it run longer to refine the detail all by itself. Training on Qwen is really messing with me. I'm breaking my own workflows and trying new stuff. The one thing I like about this run was the results remained consistent and didn't result in major artifacts, but I hated how the fur turned out throughout the run. The only thing that went wrong was the artstyle. This is pretty much unstable and needs more work but it works. Use the bf16 model because fp8 has problems with grid patterns and reduced details on anthro characters. I was wondering why my lora was bad when training was done, found out later it was the fp8 model.

Made with Qwen2512

AI Toolkit

rank 32

training resolution 512, 768, 1024

learning rate 0.0001

weighted

balanced