Krea 2
was trained with musimi tuner in 1536x1536 bucketing enabled. My poor 32GB VRAM could not handle the FP16 Basemodel - so FP8 it was. I added the final Epoch 16 as well as Epoch13 (which i think is the best one) but i do not want to waste the 6h training.
I'll try with 1024x1024 and FP16 next time - after the next heatwave in Europe...
Best results with 0.7 (maybe overtrained - but somehow the style is very versatile)
SFW works great, NSFW is strange (tongues in vaginas, anal non existant - maybe because of the NSFW captioning). Krea 2 Conditioning Control node for comfyui recommended to play around - soetimes it helps sometimes it has the opposite effect.
________________________________________________________________________________________________
ANIMA
As promised I'll release the Loras I trained for my ANIMA Model.
Nothing special here. Trained with ~250 images, mostly of hentai scenes - be careful: this turns every ANIMA Checkpoint into a cum dripping hentai mess...
Resolution was 1536. Captioning with a custom Qwen 3 VL setup and pure natural language to describe the artstyle of the images. When Qwen VL recognized a female character it always chose woman or young woman - so keep this in mind when prompting.
As it is based on ANIMA - no commercial use. I wonder how it is with on site generation (you pay Buzz).
Description
same dataset but now with WD14 Booru tagging instead of natural language and I added a trigger "@h3n741drms" which makes it easier when using multiple LoRAs or other @xyz.
Got the style better and captioning was done in a few seconds. Maybe I'll try this on my other datasets too. Recommendations are welcome. I use Anima standalone trainer an my GPU only uses ~370W instead of the 575W.
Lowerd Batchsize from 4 to 1 and LR accordingly, took nearly the same time.