Based on my Centerfold Flux 5, trying to merge in Chroma and Krea.
Diffuser version 2.3 https://huggingface.co/Tiwaz/CenKreChro_V2_3
For version 2.3 I will add the quantizations under the same "tab".
Big thanks to Spooknik for making the SVDQ versions https://huggingface.co/spooknik/CenKreChro-SVDQ
Q4 KM works fine in Comfy UI but not in Forge for example.
If you need the diffuser version: https://huggingface.co/Tiwaz/CenKreChro (not 100% tested yet).
You can try using it as low as 12 steps even if I prefer 30.
Description
FAQ
Comments (26)
This is the BEST Krea model for NSFW at the moment without any LORA, even though it's not advertised as such. Others, even with the NSFW label, didn't come close. Now one of my favorites, thank you!
Happy that you like it.
Hehe I nearly never says it is for NSFW, but me saying it contains some Chroma hint at that much 😆
The balance however was a though one as Flux tends to have a “poking nipple issue” if you push things to far.
TiwazM
Oh, you're being coy, your model can even handle porn. I tested it, but I won't post it—I wouldn't want to see an overwhelming amount of porn in the gallery here.
Thanks again for this model.
Missing LoRA as i understood just a little, the tiniest bit.
Is it better than Fluxmania?
What are the recommendations for settings?
I am not the author, I don't know if he looks in.
But generally speaking,
10-30 (max) steps:
PR: euler + normal/beta/ddim/sgm
Semi-real: euler + kl_optimal
NPR: euler + karras"
In general, the sampler is not that important; the scheduler is more important
@CG_Anahoret Thanks yes pretty much.
I tend to use 20-30 steps.
Euler simple or beta.
I heard KL_optimal can be good but still did not really try it.
Also I use no special clip or anything, basic clip L plus T5.
@TiwazM
u created MASTERPIECE,; you should see the works from others, they are perfect
We're testing it out and simply enjoying the experience.
If only you could see what gretday185 does with the prompts, but he is shy
I don't know which model prevails, but overall - none, and we are discussing this; it won't even come close to any qualitative leap.
Your model turned out better than the basic Krea, anyone who tests it will say so; I promote it everywhere.
@CG_Anahoret Yes the basic Krea is cool but does some things I do not like that much so I tried to change it.
ddim + karras
CFG 0,7 - 1,2
Steps: 14-18
My choice... today
@TiwazM Can you tell me what kind of mixing was done as a percentage of the weights of the models?
@soyv4 no as I did not take notes, plus it was a lot of back and forth with merging different blocks. I always work in a very iterative way with some times a huge amount of different test merges to get where I want.
Does anyone use clownshark? I'm interested in the optimal settings for it
@modzhahead158
Understand the principle.
The scheduler is responsible for different images look.
The sampler has little influence—you can choose any that suits you in terms of image quality, meaning only slight differences in sharpness and so on,not global details.
Only three types of schedulers differ significantly:
1)Any (Simple... Beta57, etc.—they all produce the same image except for the following two)
2)Kl_optimal
3)Karras
I personally prefer kl_optimal for almost everything because it works wonderfully with light, especially volumetric light, and light is everything for an image.
As for samplers, ipndm_v is good, and heunpp2 is probably the best (though it’s slow, it likely gives the best quality result).
this thing is super fun, great balance of models. Krea dev makes a great refiner for it.
Or you can use slightly more compact formats... I have only 12+8G videos, so it's very difficult to use such models......
Yeah, NF4 quant would be appreciated.
I’m will have a look see how it reacts, sometimes quants can mess up things.
I will test NF and some Q4 options.
@TiwazM thanks for your comment!
@mphobbit @beriaussr617 I did add a Q4 K M which seems ok, I did try an NF4 but that seemed to have more issues and was not much smaller.
@TiwazM NF4 hasn't be smaller than Q4 gguf, it's just another type of quantization. I'm not techinical specialist, but in my personal case NF4 generates 1024x1024 and 1024x1152 faster than Q4 (don't spend time for unggufing).
I don't insist much on NF4, it was just a suggestion, if it's possible.
@mphobbit I can upload one if civitai feels like it, which is a thing these days, I just thought it had more artifacts than the Q4.
I would prefer trying to make an nunchaku versions but that is another story.
@TiwazM OK, thanks.
@TiwazM I'd love a nunchaku version. It seems like they provide an example config for Flux Dev, which shouldn't be too far off for what this model needs. Needs a lot of VRAM though (H100 maybe H200). It's on my todo to figure out how to make SVDQuants.
https://github.com/nunchaku-tech/deepcompressor/tree/main/examples/diffusion/configs/model