KREA2 FINALLY: Clean dataset with watermarks removed. Currently testing the best checkpoint and strength. So far, 0.9–1.05 looks promising for style reproduction.
KREA2 NEWEST(1/2): Watermark removed, clean dataset. Sweet spot 0.9-1.05 if you coping style
KREA2 Legacy:
Use the LoRA at a strength between 0.8 and 1.7. The sweet spot is usually around 1.4–1.5.
Unfortunately, I tried to train a better version, but I still haven’t managed to get it where I want it yet, so for now this is the best version I can provide.
I’m basically pushing my RTX 3090 as hard as I can to train the LoRA in a reasonable amount of time.
ANIMA LegacyBaseV1: Just use @cutesexyrobutts
Description
FAQ
Comments (4)
Brooo this lora is so Peak, You're F#king genius at Lora training! What's train settings youre use on lora? Do you train it on civitai trainer or standalone Pc??? May i eeem uh oh... see your dataset captions👉👈??? So peak lora dude wow yea so peak ye
https://civitai.red/models/2536147/greg-rutkowski-style-anima
I train locally on my PC with Ostris AI Toolkit. I use Gemma 31B fp8 to generate natural-language captions. The dataset still has some junk and watermarked images in it, so I may clean it up and retrain later.
I’ve figured out training for Anima. Now I’m trying to work out how to train LoRAs for Krea2 properly. Once I understand the right approach, I’ll start gradually recreating different artists’ styles.
bro its just rofl on anima style lora picture chill