LoRa extract of 4th tail model.
2 sizes, XL/L and XL-alt. The bigger - the more accurate.
New version was made with some new approach, it should look better and represent more, but compatibility issues may occur. Please report if you meet any, will try to fix it.
XL-alt looks a bit different in terms of style but works best to get full color and brightness.
Before using - check out merge of full weights. May be you will find it more satisfying.
Original model comes with 8000+ artists styles, lots of anime characters (kemonomimi, gacha, touhous, vtubers, etc.), concepts and some other features.
After LoRa compression some artist are fine, some are damped but still works, others barely recognizable. Same for character knowledge, tags, style mixing. Quality tags seems working.
But who cares if you only want to use a small part that satisfies you directly on your favourite model?
Also consider using higher LoRa weight.
Works with PonyDiffusion finetunes/mixes including Autism.
Prompting:
by ARTISTNAMEby is mandatory, then quality tags, your prompt with full character name like on boorus. Check out 4th tail description for more.
Description
Medium
FAQ
Comments (8)
Thank you for sharing this! Specially the larger dim! ❤️
I've tried to do myself a lyco extract of it yesterday, but kohya kept failing when extracting the text encoders for some reason. Not sure if it's a lyco thing, but lora worked out.
Thank you for this lora extract, it works really well! What settings did you use to finetune the pony model? I was thinking of trying this method of finetuning and extract and then compare to a normal lora
kohya ss sdxl_train.py without db classes (--in_json, picture order matters), adamw (full), typical lr for sdxl. The main problem here - vram requirements, unlike lora it would be uneasy to train on desktop gpu.
Can you please explain how did you extract the lora? what did you used to compare the models?
Kohya-ss, networks/extract_lora_from_models.py
sd webui grids for comparison
@Minthybasis thank you thank you
Could you do an extract of the newest version?
Sure, today or tomorrow