Flux Klein FP8
Description
FAQ
Comments (34)
Could you post their actual outputs instead of the same image? Civitai mods have been removing models due to this for reason looking like "The showcase images/videos were not made with this model" :(
And also let's not forget that generation data matters too, so most can have a starting point on how to actually make one by themselves.
they action these stupid images??? FINALLY. I absolutely HATE these stupid cover images that have nothing to do with the model. I'm going to report everything I see.
@QualityControl Civitai Police ?
@denrakeiw yes. stop posting these pointless cover images.
@QualityControl As soon as you stop posting boobs here on Civitai, I promise I’ll stop too… dead serious… you have my word of honor.
@denrakeiw ok. pay up. look at my last post timestamp. was ages ago. I have 'stopped' .
Почему указываете Flux.2 если это Flux.1?
Потому, что авторы Flux2 выкатили новую модель - Klein на базе Flux2
это не flux.1
Huh? That’s flux-2-klein though 😉
Yeah, but creating an FP8 model takes under 5 minutes on my 5090.
For NVFP4 I’d need to rent a server, and that takes days.
That’s why there are barely any people doing NVFP4 models as a hobby.
NVFP4 is more difficult to make, and quality may end up very low.
@denrakeiw BFL has fp8 scaled models on HuggingFace, you don't have to convert
@qek Oh nice, I just noticed it now. Even nvfp4 models are on Hugging Face.
@denrakeiw I don't really understand. nvfp, even 9b, in 4 steps (cfg:1) generates approximately 5 seconds on my 5070. The model takes about 5 GB. the quality is very good.
@saymannsk330 nvfp4*, I heard it needs Comfy Kitchen and CUDA 13
@saymannsk330 Not everyone works in low res ;) An 8K render takes 17 minutes on my 5090 using a combo of 9B and 4B models...
They're only faster on 50-series cards. For others, they're only interesting because of their small size.
@denrakeiw Flux2.Klein models generate up to 4 Megapixel. Anything higher should be done via upscaling. Furthermore, the distilled model generates in 4-8 steps; only the base model uses more steps.
@sizyking Unfortunately, artifacts often appear at 4 MP — I think the sweet spot is around 2 MP, and then using divide & conquer to blow it up to 4–8K.
Can you upload the full BF16/FP16 model too?
It's on huggingface
@denrakeiw I know, but it could be used here for inference and to post images in the gallery.
@mmdd2543 For licensing reasons, it can’t be used here for inference — at least not if I upload it myself. If Civit pays for the license, then a Civit employee will surely upload it and enable it in the generator.
@denrakeiw Ok, but still, why is uploading an 8FP version allowed but the BF/FP16 is not? Not trying to argue, just curious.
@mmdd2543 different copyright licenses thats why. its really dumb.
@mmdd2543 U can upload it but can't use 9b in the gennerator
Why does this model show up among the Z-Image Base models?
This 9b base model gave odd output with comfy. Downloaded full base straight from Black Forest Labs and all is well. Nano Banana meets SDXL. Perfect for 24gb cards. Faster than Flux1. LORA training takes longer but well worth the wait. The capabilities of this model are truly amazing on consumer hardware. Models like this are why all serious computing will be forced to be done in the cloud one day.
I'm using a klein-9b-fp8, it's a great model. But, I'm having trouble with cropping the composition. The model slightly crops and zooms in on the original image, making it difficult to match the source image with the generated image (even if I take it right before the sampler). It's incredibly annoying and this didn't happen with FLUX Kontext if i get image right before sampler. If anyone could suggest a solution, that would be great. I tried adjusting the workflow as the FLUX Kontext workflow, and I've tried everything, but it seems like there's some kind of bug in the architecture or something.
The detail is better than in Zit, but it's impossible to use because of the anatomical deformities. It's like I launched SD1.5 on steroids. In most of the photos, there are either three hands or four fingers.
You're doing something wrong. Impossible to know what from what you've said but I don't have that problem at all.