09/07/2026 - Now added reference image edit mode to my workflow.
I am a bit late to the party on Flux Klein models as I have been trying out other stuff, but I am back to see how far I can push Black Forest Labs most accessible model.
Anatomy can still be hit or miss on the roll of the dice (usually arms/hands missing)
8-12 Steps of Euler/Simple seems to be good. (To get better coherence/anatomy than 4 steps, but you can get lucky with 4 steps and enough gens.)
Description
09/07/2026 - now added reference image edit mode.
Initial Release
FAQ
Comments (19)
I'm guessing you trained on a bunch of GPT Images 2 outputs (not exclusively, obviously), because I can see its artifacts on many of your examples. Is that right?
No. I did notice some strange noise on some images when testing earlier, when I was asking for "film grain", but that also appears in the Base Flux Klein model. I think that is because Flux Klein is a distilled model which increases speed but comes with quailty downsides.
@J1B Hmm, interesting. I see that your images are quite high res, so presumably they are upscaled, so maybe it's that? There's a grid pattern in the woman's top in the one with the woman with headphones lying on the bed, that can be seen in the full res. That's not normal for Klein.
Do you have plans for the fp8 version?
The fp8 version is there now: https://civitai.red/api/download/models/2948023?type=Model&format=SafeTensor&size=pruned&fp=fp8
Seems like it might be a little less consistent with coherence/anatomy, but can look almost as good: https://civitai.red/posts/28621486
Hey! Great samples. What does this checkpoint do? Does it just make klein more realistic? Does this affect the anime or other styles?
It's mainly adding NSFW anatomy cababilites and concepts while leaning more towards realism and a slightly more noisy amateur phone photo look. I will add some comparison testing images I have done.
I haven't done a lot of testing with Anime images, but it seems this model can pull some images more towards realism if it is a semi realistic style (which is mainly what my models are know for) but it still seems surprisingly flexible.
Do LORAs work with this great checkpoint? Also do you prefer Flux Klein or Ernie image?
Yes it works well with Lora's as it is not too far from the base model, just better NSFW. I think I am leaning towards liking my Klein model more than my Ernie, as it looks pretty realistic and Ernie looks a bit more plastic, but they can both have issues making correct anatomy consistency, they are quite similar , but also Klein has more loras right now.
@J1B how do improve anatomy as I tried the anatomy fixer Lora and slightly higher steps but anatomy is still hit and miss. I was wondering if I can use Flux Guidance or something like that but I think it's more a model thing as anatomy is a problem with most Klein checkpoints...
Very good model to convert photos to drawings or any traditional art.
Hi, I've been testing this checkpoint and it's great for following instructions, especially if I increase the CFG to, for example, 2. The problem is that it doesn't preserve identities or hair when using image editing, or at least I don't know how to fix it. Are there any methods or suggestions?
Thanks, I have had some good success with face insertions in editing mode, it does seem to depend on the face and sometimes doing a few rolls (10-12) to pick once image with the best likeness. I will upload my updated workflow with reference image inserting.
@J1B Hi, oh, great if you share, thanks for replying
This is awesome, you managed to keep it real solid. Good job and thanks!! Hope you keep making them.
This checkpoint uses Flux RMSNorm keys ending in .weight, while older ComfyUI’s Flux/Flux2 detector expects .scale. You may need to upgrade ComfyUI to support loading this checkpoint.
Ah thanks, I think I have had a few people telling me they just get black images out, but I haven't been able to replicate it.
Is there a very good updated version?