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!
Miraclein focuses on photorealism and more grounded/grungy images. Not amateur style, but not fantastical either.
Generation Settings
Most settings work, but I use:
12 steps
Euler
CFG 1.1
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...
Editing
It works fantastically! Try it out as an edit model!
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 (46)
So the BF16 version of this model has instructions to use the FP8 Qwen text encoder on Forge Neo. That can't be right, right?!
The main description is for ALL models. So just use whatever you need to use.
@6tZ Wouldn't it have been easier and more helpful to others to say, "yes FP8 text encoder is not going to work on BF16?" Seems odd you have taken an adversarial approach to those trying to call out where improvements should be made.
@dnmiela772 It's not a question I know the answer to. I don't use Forge Neo, so I don't know how to answer it.
And I just actually tried to be helpful by stating that the main description is for all models, so it may not be relevant to this specific question.
@dnmiela772 If your GPU supports it, you can use any BF16 / FP16 / FP8 / NF4 combination for models and text encoderon ForgeNeo
2.0pf8的一致性没有1.0的好
Will the flux 2 LoRa also work on this model?? Or how we need to train LoRa on this model? If so, then how we train this model LoRa?
so I know it's NSFW, but having a hard time keeping breasts from being exposed when I want them covered, like a shirt, or sundress. Any tricks to be more modest before the real NSFW fun begins?
If you want it to NOT be exposed, do not mention it in the prompt. You can always use a lora for size and shape but that's what I've found works.
Not sure what specific part you are referring to tho 🤣
@Walnutphysics The model is definitely a bit overfit, gonna try to dial i back a bit for the next version.
@Walnutphysics you answered perfectly! What breast lora do you use for this checkpoint?
For Flux2 with these models, does it matter if we use an uncensored text encoder?
It's more of a variation than an improvement. It doesn't help with any NSFW concepts.
Was this model trained on any anal penetration? If so, any tips on getting it to ... ummm... perform! lol
I like the style of this model. But when I try to dress up the characters, it always leaves them naked. That's a bit of a shame.
These are not the clothes you're looking for.
It appears that Kleinova v3.0 BF16 and Miraclein v3.0 BF16 share the same hash
What the hell! I've fixed this twice already! What's going on Civit! :|
@shoes22 can you confirm that you downloaded it recently?
This was fixed a few weeks ago. And if you downloaded before then, your old version would be different.
I just re-downloaded them from civit, and I confirm THEY ARE DIFFERENT. So I bet you have the old versions.
File: miracleinNSFWGeneration_30Bf16Fp8.safetensors
CRC-32: 418b17dd
MD4: 21f75e2be75ff175aad76b957932a591
MD5: 0b98e0e67c9ca198b09b3f58e19a51d5
SHA-1: 897c6859d60f880ae56c5512422c1132e438f9df
/////////////////////////
File: kleinovaNSFWGeneration_30BF16FP8Bad.safetensors
CRC-32: 43d9446e
MD4: 3b838ae4b908c346fdbdc2ce010762f8
MD5: 0ebf81567d3fe0dee42e23675be4c706
SHA-1: 7fea3f0a4b3b50228dea9e9065b0ae58fd0db8d4
Anyone able to post a good working img2img workflow?
In my tests in Forge Neo, none FLUX KLEIN works on Img2img. Didnt check on COMFY
Kleinova has edit examples in 1.0 FP8:
https://civitai.red/models/2547526/kleinova-nsfw-generation-and-edit-flux2klein?modelVersionId=2863261
use the flux klein suggested edit workflow here https://docs.comfy.org/tutorials/flux/flux-2-klein
that worked well for me
I've just uploaded one, hope can be useful
@marcomentre792 where?
@s5miyagoo589 here https://civitai.red/user/marcomentre792/models I uploaded an image 2 image workflow with 3 separate image reference and (just updated) a workflow for Pulid with klein that now works beautifully :-)
Won't the generation speed slow down when using "12 steps"?
Slow down compared to what? 12 steps is pretty fast usually. It's all relative I guess?
@6tZ If it's just a simple modification and editing of the image, will the default setting of "4 steps "reduce the generation quality?
@TTTTTVTTTTT Yes, try for yourself though. It's not worth asking others for their opinion when you can make the test yourself :)
I find 12 steps a minimum for my quality threshold, but that's up to each person.
With less than 12, I find that it may be confident in a limb that shouldn't be there. i.e. you get extra arms, or replace a penis with an arm. It basically "rushes" through the early shaping phase which is important.
@6tZ I noticed that "KleiNova" is also a work you released, and you marked V3 as "poor" in it. The latest one is V2.2. I couldn't find any difference between "KleiNova" and "miraclein" in the introduction interface either. Could you please explain the difference between the two?I have downloaded and used both of these checkpoints, but I haven't noticed any obvious differences.
@TTTTTVTTTTT They are different models trained on different things.
that was best nsfw model for editing, it doesn't even need masking. thanks
Hello creator, int8 versions have recently become popular; would it be possible for you to release an int8 version of your "F2K-9b-miracleinNSFWGeneration_30" model?
Not quite sure how to convert to this :|
@6tZ F2K-9b-miracleinNSFWGeneration_30_fp8真的很好用，哈哈，如果能有int8_convrot.safetensors版本的模型就更好了，这种速度会更快一些。
wow i ran it all this time with 4 steps and it was already amazing i se recommended is 12 steps?
I never found 4 steps to be good enough for any model. It CAN output good things, but you get too many failure cases :)
This does not work with wan2gp gives missing key errors
Missing keys: ['double_blocks.0.img_attn.norm.query_norm.scale', 'double_blocks.0.img_attn.norm.key_norm.scale', 'double_blocks.0.txt_attn.norm.query_norm.scale',
etc.
Excellent! Thank you for your work :-)
Hi, thank you for your hard work!
There will be a 3.0 with nvfp4? that would be awesome :D
I don't know how to convert to it. The first version was using this as a base, but then I re-trained for BF16. Any ideas?
This model would be perfect, but it's too biased to women being in certain poses & positions.
When you remove clothes from a woman laying on her side with her legs together, it will still generate an entire vagina on her pubic mound/lower abdomen (which makes no anatomical sense).
what is the difference between this model and your other model KleiNova NSFW [Generation & Edit] [Flux2Klein]
The difference is that they are different models. Trained on different data, and thus, resulting in different results.