This is the preliminary Krea test of SabuArt. It has a few shortcomings, chiefly, like Z-Image, the base model will have trouble meet you halfway in places.
Second, unlike Z-image, and Anima, Krea doesn't seem to have picked up on my tenuous attempts to breakdown the artwork into different coloring styles
Otherwise, the trigger is 'in the style of SabuArt-Sabudenego' Krea will make it mostly in his muted, monochromatic style, but depending on what else you use, and what checkpoint, you might have a different experience.
Description
FAQ
Comments (8)
Nice work. What do you think about krea 2 ? the images looks good, is it better than others for you?
I'm a little on the fence right now. It's just too new to me. I think the outputs are really good, and it follows prompts pretty well, but it's like Z-Image in that if something isn't in the base dateset you're pretty much out of luck. Using loras to get around it works, but the best one I've used so far loves adding fluids to everyone.
It's difficult for me to choose between Krea or Anima, and I kind of want to test Krea against ZIT now to see how they compare.
@Gwenwithgreeneyes What kinds of concepts are giving you trouble? I found that you can do a lot if you just word your prompt precise enough.
@leerlauf Mostly it's the explicit stuff, with a NSFW lora it will usually do alright, but it's never going to be as robust as with a model that's trained on it.
That, and I get the feeling that Krea.2 is much more realism ready. Without a lora telling to to make something in a artist style, or comic style, seems pretty limited.
But this could all just be inexperience for me right now, like you say.
@Gwenwithgreeneyes You can get a whole bunch of super different 2D styles right out of the bat. I am using the NO8D-prompt-library node to make this more convenient, but all that really does is copy-paste a pre-written text into your prompt. You can just manually copy the prompts instead from the spreadsheet files: https://github.com/no8d/ComfyUI-NO8D-controls/tree/main/data/krea_style_libraries (alternatively look here to see the prompts in action and to get a google drive link so you don't have to download anything: https://www.reddit.com/r/StableDiffusion/comments/1uzdj7o/krea_2_styles_wildcards_txt/)- you can also take a look at the official moodboards and copy or combine the different prompts from there: https://www.krea.ai/app?gallery=moodboards
About NSFW - I would recommend an NSFW-ready checkpoint over one of the nsfw-loras. I'm getting better results with those, although they do of course influence and alter the style. Currently I am using lustify for quick testing, and Sick Ollie for the more high quality upscaled end image. Just general erotica and porn works fine, but Krea does indeed struggle with the more fetish-related content.
@Gwenwithgreeneyes It seems definitely better than ZIT and anima to me. Like you said, it follows prompts very well too. My only problem with it, it doesn't support nsfw naturally so you need to use nsfw loras which some of are pretty heavy and slow down the generation. Otherwise the trained loras, characters etc are also really good and accurate compared to any other models. Krea 2 seems very nice.
@leerlauf Thank you for these resources! I'll get stuck in later and see what I can manage. I'll need to experiment more with different checkpoints and loras, that's for sure. Thank you Leerlauf, I adore what you've posted so far 🤗
@chrisss1
I have to admit, putting in old ZIT prompts into Krea2 has made them better, and even when the lack of NSFW makes it difficult to make this pose or that, the output is still good enough that I don't mind it's off the mark.