Krea2Raw v2 released with a much Larger Dataset 472 Images
Trigger = dcd@rkf@ntsy.
Trained @7000 Steps, Rank 32 with CivitAi, Captioned with GPT4.1, Will be slightly different style to v1 and 0.5 with more reds and blues and a little less painterly. Check my profile gallery to see more examples or at the bottom of lora page.
v1 & v0.5 Trigger Word = dcdrkfntsy.
Krea 2 Raw v1 released and is trained at 5000 Steps it is closer to the intended outputs and a bit more painterly, certain faces will show yes but you can combat these krea releases by droppping lora weight to .7-.6 and you should see a lot more variety! this style was a test with such a little dataset to see what 22 image datasets can achieve.
Krea 2 Raw 0.5 was trained at only 2000 Steps to accomodate people who were having a hard time with the last releases but it is a little less painterly.
[ ZImageBase, ZimageTurbo, KLEIN4B] Trained with AI-Toolkit for 5750 steps [Klein9B] for 7000 steps Rank 32 Tagged with very basic captions consisting of 20-30 words with JoyCaption Beta One (22 Images Total) - Probably a little overcooked (Needed to get the high steps to get the aesthetic because im an amateur at making loras lol! if your having trouble getting what you want drop weight from 1.0 to .75.
All images have generation details/ Workflow for ComfyUi just click the image you want, then save, then drop into ComfyUI or Open the image with notepad on pc and you can search all the metadata there.
If you use this lora in merges or upload to other generation sites please mention my civitai page on your post, Thankyou
https://civarchive.com/user/dkpc69
Description
FAQ
Comments (8)
How do you feel about Z-Image Base for art styles now that you've trained on it?
It handles scenes much better than turbo which im figuring out "turbo seems to struggle with rocks and grass - maybe it is the way i train (But for this specific lora turbo looks alot better than base which is weird! the variety of base is much better than turbo, for photo realism and character loras i prefer turbo - Overall i prefer flux Klein 9b over all of them, Only negative a i have with klein is the extra limbs of people it sometimes generates but that could be down to my bad prompting and/or captioning
They are all pretty easy to train and roughly the exact same standard settings i use for each lora - base and turbo look really bad when you see the samples as it trains but when you use them in comfyui to generate images they come out alot better
and this will probably be the on klein4b model i will train i don't really like the outputs, Unlesss i find a good setting in the future
@dkpc69 "base and turbo look really bad when you see the samples as it trains but when you use them in comfyui to generate images they come out alot better"
Shit... makes me wonder if all those weeks I wasted trying to train it weren't actually wasted. I was only getting bad results no matter what I tweaked. I guess I should try again. Thanks for that info.
@Jellai yeah definitely try them out! Even if the samples look half good they’ll look a lot better in comfy so don’t take it as the Lora isn’t training very good! I’m still learning too they’re only just some of the things I’ve noticed after training 100 Lora’s
0.75 is way too much, it repeatedly brings up the female face from one of your sample images. It still works around 0.3-0.4, even without the trigger word.
One of the very best painterly effects there is, but unfortunately in ZIT has a devastating effect on character proportions, especially legs and waist, and turns all clothes to "ridiculous". Also brings the immediate beckground in to the foreground. A real shame, as a standalone media lora it would be fantastic. Performs better in ZIB but is not as pretty.