unStable Dissolution KREA2 Turbo
LoRA Version of this Model: KREA2 Turbo NSFW LoRA
For all those who want better NSFW content
Sample Image Settings
Tool: Forge Neo 2.26
Steps: 8
Sampler: Euler
Schedule type: Simple
CFG scale: 1
Size: 896x1152
Model: This
Module 1: wan_2.1_vae
Module 2: qwen3vl_4b_fp8_scaled
Diffusion in Low Bits: Automatic (fp16 LoRA)
FREE / No Scam for BUZZ like other people do by merging another vae or encoder to the same base model and call it AIO or push "updates" every 3 days to start a new early access for another 5-10k buzz
No bypass, trained NSFW
No Image edit, real Outputs (see Metadata)
Description
FAQ
Comments (37)
Sample images look good. If you plan to make gguf-int8-fp8 versions, I've been reading lately that INT8 ConvRot versions of models kicks some serious butts in speed and quality...
Thanks
Do you have any comparisons? @lolmao500
fp8 would be really cool. i can't run models this size.
Like most users. if this model gets some good reactions, i will do the fp8 and gguf. share it. :) thanks for your comment!
@Peli86 +1 for fp8 from another VRAMlet here!
https://github.com/silveroxides/convert_to_quant you can easily convert to fp8, int8. I suggest int8. It only takes about 3 minutes to convert on a 16gb card.
@Peli86 How can we give you good reactions if you don't publish models we can use and then remove people's comments too. I converted both to FP8 scaled and MXFP8 spending over an hour after downloading it all night. Then guess..., both of them are highly censored. I thought this was an NSFW model it sure doesn't look like it or the FP8 conversion didn't work. You got a well deserved downvote, Peli.
@ferrrett33 1.: I'm not a mod, so I can't delete anything.
2.: You're converting a model without the necessary files that make the change possible. What you converted was 100% the krea2 raw model, not my model. You can't convert that.
3.: When my FP8 and other models are released, you can apologize.
@Peli86 Well great thanks for telling me that now, I don't know how FP8 conversion works with Krea 2, that's why we shouldn't be doing this. Are you aware consumer video cards cannot handle this? Oops, I mean 98% of them can't.
I quote a chatbot: less than 2% of PC gamers own a graphics card that originally retailed for $1,000 or more.
I just got a brand new 5070 so I'm all good to run this, everyone else's that is, except yours because you're literally the only one right now who doesn't publish in Fp8.
Whatever bro, have fun, enjoy life.
@ferrrett33 I love these appreciative comments on FREE content, thank you so much! 🖕
@hempires published
Always fan of your work. The Lora Version works flawlessly.
May you provide an fp8 alternative? i have limited vram to use bf16 properly.
22G?
Expected more?
as a mac user we appreciate bf16 a lot!
@sanihands11225 Yes great at a quarter the speed.
int8: I converted this to int8 but couldn't get mxfp8 to work. Could you possibly make a mxfp8 version for 5000 series cards?
never converted anything to int8 or mxfp. fp8 will be released.
use this to convert to mxfp8/nvfp4 or int8, i use it all the time for mxfp8 conversion. https://github.com/silveroxides/convert_to_quant works perfectly with z-image,anima, flux klein, ernie and even wan and ltx. it does support krea2 also but have not personally tested any krea2 model yet. just install and use the commands its simple, use the simple option for fast conversion without extra tuning, not needed for most mxfp8 minimal difference and you save 30-40 minutes. 30-40min without simple mode. 30sec - 1min with simple for full model conversion. install and then your run a simple command. : convert_to_quant -i krea2_model.safetensors --mxfp8 --comfy_quant --simple --krea2 -o krea2_model_mxfp8_convert.safetensors
Hi! The model looks very nice! Can i run it on Comfy?
Yes you can. I have been using it on Comfy since release
@WikkedAI Great! Could you please give a recomended settings for the sampler?
@stnlnxwll666 i have a simple workflow on my page you can try
@WikkedAI i checked your WF, it's awesome! thanks again. One question: in your notes, you recommend several loras.
Krea2 NSFW v3
Realism Engine Ideogram Krea 2
krea2filterbypass
Do they need to be alternated or switched on at the same time? I've tried both ways, with a power of 1 each, and I get interesting results. But I can't figure out if the quality really changes with the simultaneous use of the above-mentioned loras or not?
@stnlnxwll666 they are mostly loras that i enjoy and use on all images. completely optional
Thank you for actually uploading a BF16!!!!!! Also to anyone wondering YES it works on Comfy. I have many images below to show its output.
They should do both this is so ridiculous. Took me 1.5 hours to dl the 24 GB file from CivitAI, after it got aborted twice too. Finally, after I let my computer run overnight just to dl this file I convert it to FP8 using GGUFY. I picked the scaled E4M3 except it looks totally censored, I think something went wrong and it's not supposed to be that way. But why do we even have to do this? Hey I feel young again like it's 1995 that I keep my PC on over night to download a single file.
I haven't noticed a quality improvement from using the BF16 versions of Krea 2 on my 3090 and it runs quite a bit slower than the FP8 which I cannot tell the difference with.
Nice model! I appreciate the buzz hungry comments also, it's annoying to see. Cheers!
I realize it's early days still for Krea 2. But z-image-turbo is still the king in my mind for ethnic diversity. Indian genotypes like Tamil, Bengali, etc. as well as complex skin textures are handled easily by z-image-turbo. I would like to see the same thing happen to Krea 2 over time. krea 2 may even have the capacity to exceed z-image-turbo over time with the right fine tunes.
Any particular ZIT model you would suggest?
@etherslayer11 Most of them are pretty good. Big Love, cyberrealistic, SickOllie, Try those.