Introducing my NSFW Krea2 turbo AIO checkpoint.
Based on mostly a LoRA trained by me in AI Toolkit, plus a few existing LoRAs with smaller weights.
AIO version, so use the Load Checkpoint node in ComfyUI.
I've used it with these config:
steps: 12-14
CFG: 1
sampler: euler_ancestral
scheduler: beta or simple
size: 2048x2048
feel free to experiment with the parameters.
Description
FAQ
Comments (36)
really helps to have prompts in your example images when posting a checkpoint or lora
This workflow didn't work, what could be the issue?
same. says it doesn't recognize model
@maint4life Manager > Update All
I get the same error "Could not detect model type", tried Manager > Update All but no luck. LLMs don't give any guidance either. ¯\_(ツ)_/¯
@NothingsPerfect Updating the ComfyUI fixed the issue.
@aftermathremixes945 It worked, thanks bro.
@sriramakumar7799888 Appreciate it, mine was stuck on v0.24 and wouldn't update, had to force install it. working now
Hello, what are the approximate percentages of females vs males in the training data, especially young attractive males? Is it like 90/10 or 80/20?
Unbelievably good. The only thing I don't get is that it very often creates a lot of wet skin/water streams even when prompting "dry skin" how can I avoid this?
i negative prompt this: water, rain, river,
@necroryona thx, but I avoid cfg >1.0 to keep generation speed moderate :-). A Wetness Slider Lora was just released. Maybe this will help. We'll see : https://civitai.red/models/2738333/wetness-slider-krea-2?modelVersionId=3079282
Please tell how many images in train dataset?
This is the best damn NSFW Checkpoint I've tried so far. Almost perfect, a 9.5 if not a 10.
great! if possible, bf16 (or fp16) would be great for all the mac users (and big vram folks), FP8 doesn't work on any MAC and the 'patch' is very slow and only half works.
Converting to bf16 isn't going to speed anything up is it? The patch is just doing the same thing the conversion would do. I converted fp8 to bf16 on my Mac to test them both and they both got the same s/it.
Very nice AIO. Would you mind sharing an openpose/depth wotkflow that would work with this model? Like taking a basic image made from a camera in blender and turning it into a new image?
this checkpoint is slept on, it can do wonderful things, can't wait for a newer version that supports actual licking, kissing direction since Kera2 sucks at those.
Dude this is really good, top league, you're awesome, how about a name for it instead of ungarian fckmaybe as the only identifier for your top end checkpoint.
This works very well, even for SFW images it will be a little bit more sexy which I like ;)
I wondered if it could be converted to ConvRot (I don't really know what it is x) ) then it could be like 40% faster like the krea2 int8 convrot version from the Comfyui repo ? I asked Claude to help with that but it didn't work lol I got only noise xD
Or maybe it's already there in the first place, and after the first load it's already pretty fast anyway
This work in NEO?
Yes, Krea is supported by neo: https://github.com/Haoming02/sd-webui-forge-classic/issues/1254
Just update your version.
@mphobbit Aw jeezus. Update as in fresh install of a newer version? I'm a big dummy when it comes to all this kind of thing with the basic setup.
@antonovfedir193 just download a fresh version ("git clone ....") and replace old files with newer copies.
@mphobbit Thanks for the info. I did that, and don't see a tab for Krea. I think I'm missing something, lol
Check yourself:
git clone https://github.com/Haoming02/sd-webui-forge-classic sd-webui-forge-neo --branch neo
Without branch you could download "classic" version.
@mphobbit Turns out I was making a mistake whilst installing. Ran into another issue, though. Z Image models work just fine, but whenever I try this model, the error
AssertionError: You do not have Qwen3 state dict!
pops up. Is there another text encoder I should be using?
@antonovfedir193 you have to use qwen3-4b text encoder for Z-image. Krea has qwen3vl-4b (they are different LLMs).
My choice of TE for Z-image is https://huggingface.co/jiangchengchengNLP/qwen3-4b-fp8-scaled
@mphobbit Ah. That's it, then. I have the exact text encoder you mentioned. Where do I find the one for Krea? Thanks so much for all your help.
Found these two pretty easy:
https://civitai.red/models/2731465/qwen3-vl-4b-abliterated-comfyui-krea-2-text-encoder-bf16-fp8
https://civitai.red/models/2728378/qwen3-vl-instruct-4b-heretic-7refusal-convrot-alternate-krea-2-text-encoder-for-comfyui
Will pick one.
a gguf would be nice for the low vram users. q4, q6
INT8 pretty please
runs well on my 6800 XT (just a bit too slow for day to day use)
It seems to give every woman a shaved pussy, no matter what the prompt is. Even if I add pubic hair, trimmed busy, anything like that, they all come out completely hairless.
i dream of a convrot int8 version of this checkpoint.