Flux 2 was not what we wanted or needed, but Klein appears to be what we deserved! Let's unlock the full NSFW potential of this model!
Kleinova focuses on professional photography and beautiful shots. It's vibrant and colorful and over the top.
Generation Settings
Most settings work, but I use:
12 steps
Euler
CFG 1.1
Editing
It works fantastically! Try it out as an edit model!
FP8 Download
For some reason, when you upload multiple formats (BF16 + FP8) on one model, which they now support, they hide the FP8 under the BF16 model group in the side menu...
GGUFs
Do it yourself:
https://github.com/qskousen/ggufy
Forge Neo
Should work fine using qwen3_8b qwen_3_8b_fp8mixed.safetensors and for VAE FLUX.2-small-decoder. Make sure to be on the correct preset.
Description
FAQ
Comments (27)
can't download fp8, file not found.
It's working now, try again.
@6tZ still same
actually it worked on 5th retry, civit issue
@dezeptup125 works, for me, now,
still happens, must be civit, i just open it with middle click
I noticed in one of your workflows you use image stitch but I can't seem to get the prompt to recognize that the image is two in one. Any tips?
IE: an image of the jacket top and person bottom.
Prompt: Put the jacket from the top image on the man in the bottom.
Hmm, not really any tips, try looking for guides online. I'm sure I've seen some official documentation of this.
@6tZ all good, really is a small thing. Great model! Thank you.
FP8 is good enough for me I paid for the Fp16 eraly access to support you. Please keep up the goodwork.
i use comfyui with --gpu_only flag but on 3090 fp8 checkpoints dont fit so i cant use loras with it.
-initial testing, i2i only
-kleinova fp8 - without using any Loras -
klein TRUE series , big love klein v2 , moody desire , flux trait , dark beast i dont remember using pmaster v3 probably bcuz at that time i didnt know how to quantise models so dont consider pmaster v1 to v3 in this testings
tldr;
currently i think without using/supplementing with any Loras KleiNova is the best 9b nsfw checkpoint.
-------------------------
NOT-so TLDR;
-excellent character consistency
( only rivalled by big love klein v2, klein v2 is bit better in that one category , )
it tries to change the pose BUT BUT BUT adding prompts like
dont change original pose keep the original pose , keep original hand positions, dont change original hand positions almost solves it. dont worry you wont be disappointed.
(tip- this can probably be easily further mitigated by using consistency loras along with it)
-knows about attires (most nsfw checkpoints fail they either dont know attire or just merge body/body-parts with attires)
-compared to other checkpoints it does better job of oppai (once again only rival is blove klein ) and does better second pair of lips (kleinova is best out of the box for second pair of lips)
-a lot of concepts already trained and built into it, looking forward to further more concepts
-great hands production btw
-other checkpoints have hard time handling hand-holding poses which this one doesnt as much.
Thank you for uploading bf16 version of this checkpoint. :)
cant wait to try it with loras.
getting "Could not detect model type" in comfyui even when using one of your images for the workflow
That's mega weird...
@6tZ yes claude said its keys or something were reading as a lora, idk, probably something wrong on my end
@jbsbiome Hmm, I guess you are certain you are using it as a model? These go into the diffusion folder of ComfyUI.
@6tZ yes correct folder with my other klein's, flux etc
@jbsbiome Ooof, yeah then I have no idea really. Sorry about the trouble :/
@jbsbiome I have the same problem, did you solve it?
@Picslover no, deleted and moved on sadly
Getting same error on 2.0, 1.0 works.
I feel like it must be a "ComfyUI Updating" issue. Either MY version is fucked, or YOUR version is. Since more people are having it, maybe it's something on my end? Unfortunately so unstable.
I can confirm that I can still generate with the model =_=
@6tZ Turns out it was a comfy update issue. Works now!
nvfp4 - very bad quality, image washed out and blured.