ATTENTION: The LoRA version is NOT a checkpoint, but a LoRA for local gen.
UPDATE2: Not much change in γ, still working toward keeping the look i like but correcting small quirks.
UPDATE: β is a less "white sticky fluids glazed" version of α 😁
μKrea2 (mu - Krea2) is my own final Krea2 mix based on several trained LoRAs (including nKrea) and some extracted LoRAs made as an experiment of hacking sd_mecha and learning how to unquantize INT8-ConvRot checkpoints (because undoing stuff is the best way to learn of to make stuff).
And since both the following events happened:
I finally found out how to convert stuff to INT8-ConvRot (via convert-to-quant),
stable-diffusion.cpp is now compatible with INT8-ConvRot,
here it is for your enjoyment!
I am releasing it as Early Access at low Buzz price to avoid immediate scraping.
Since this is Krea2 checkpoint, nothing special on usage, here is just some info on how i use it (for the showcase pictures for example):
ER_SDE/Smoothstep (aka Beta)
CFG 1 - 8 Steps
VAE: HDR VAE by @Felldude (basically the best FP32 VAE for WAN2.1/ANIMA/Krea2/Qwen Image)
All pictures made using stable-diffusion.cpp version master-874-656a135-2-g2ea8aff, commit 2ea8aff, compiled from source on ubuntu 26.04 with CUDA 13.2 with an RTX 3060 - 12GB. Each 960x1440 px pictures generated in around 60s (including decode).
Command line used (to start server):
./sd-server -l 0.0.0.0 -p "" -n " " --sampling-method er_sde --scheduler smoothstep --steps 8 -W 960 -H 1440 -b 1 --cfg-scale 1 -s -1 --clip-skip -1 -t 0 --rng cuda --sampler-rng cuda --diffusion-model muKrea2-INT8-convrot.safetensors --vae finalcut_vae.safetensors --llm Qwen3VL-4B-Instruct-Q8_0.gguf --vae-tile-overlap 0.5 --vae-tile-size 32x32 --offload-to-cpu --vae-tiling --fa --eager-loadAll pictures batched using a small custom sd-client that talk to the server using the API endpoint (i'll release it if someone wants it).
./sd-client --help
Usage: sd-client [-s "832x1216"] -o output.png -p "A picture of a cat"Thanks for reading!
Special thanks (for the great pictures they make that helped me find a direction for this):
@llikswonskcalb
@ElectricDreams
@Chumple
@lonecatone23
@Sophorium
@daceheg192491
@Liant
Description
WARNING: THIS IS A LORA, NOT A CHECKPOINT
Posted on Civitai, upload found elsewhere are FORBIDDEN and done without the author permission.
FAQ
Comments (23)
esta brutal . Una pregunta . Podrías llegar a crear algo así de realista para illus? es espectacular , pero no creo que sea posible , verdad ? illus ya quedo como pony XD
Desculpa mi espanol! Illus y Pony estan Unet y CLIP modelas, ellas no pueden crear detalles complejos.
(In case my Spanish is just gibberish: Illus and Pony use a Unet and CLiP architecture, they can't output too much details. More modern DiT architecture are better at this)
@n_Arno Thanks for the effort in writing it in Spanish. It turned out pretty well XD.
I understand what you're saying, and it's a shame that Illus has fallen behind these models. In the end, everything evolves, and you have to keep up to avoid being left behind.
Thanks for the reply. <3
The lora version is also amazing. Thanks for sharing.
bf18 would be great
bf18? I suppose it's BF16, sadly... uploading ~24GB is a pain in the a**, so, no won't happen. All tools can now do INT8_Convrot and the quality is almost equal to BF16, no need to be bother to do so ^^;
(And if a BF16 version is wanted to merge it or whatever, reverting from INT8_convrot to BF16 is possible, just documented nowhere, but since i figured it out, i'll suppose other people can too :D)
A snippet of code regarding this:
```
gs = jload(bytes(scales[key.replace(".weight", ".comfy_quant")].numpy()))["convrot_groupsize"]
return rotate_weight(weight * scales[key.replace(".weight", ".weight_scale")], build_hadamard(gs, dtype=torch.float32), gs)
```
@n_Arno sorry, of course BF16. Guess uploading is pain. So is downloading and storing these monster files. But I'm on Mac and all I got to run is BF16 unfortunately.
@user081525 Ah, then a GGUF version should be ok for Mac i think, what are you using? Comfy?
@user081525 you can also get the LoRA version and use it with base Krea2 :D
@n_Arno using ComfyApp, yes. Did struggle wit GGUF also. But will try out the LORA version, thanks!
@user081525 It is "marked" as a checkpoint here since it's a version of a full model, but it is really a LoRA which goes in your LoRA folder and should be loaded with a model as usual
@n_Arno if I use the LORA version what checkpoint would you recommend
@user081525 Base Krea2 Turbo. The LoRA was made by extracting the difference between the full model and thid base, on other uncensored Krea2 models, it may be too much 😅
How do I avoid the women always being covered in cum?
Yes... so, it is indeed a shortcoming of this version, my bad! ^^;
You can leverage the LoRA version and use lower strength or use CFG ~1.5 and add "wet skin" in the negative.
You can also use the LoRA version on top of the checkpoint with a small negative strength (like -0.1 ~ -0.2 max) to lower the effect ^^
Try using this at -2 strength. Works well for me.
https://civitai.red/models/2708727/cum-slider-lora?modelVersionId=3161687
@gameboynitro oooh, nice trick, thanks!
@n_Arno no problem, nice model!
edit: Also if you combine with https://civitai.red/models/2738333/wetness-slider-krea2-zit it removes even the water drops!
@gameboynitro Yes, i did a test and this one does not alter at all the pictures 👌
2 problem,1 auto making cum sign, 2, tongue change like a dick.... and good checkpoint
Oh yes, i KNOW it is far too horny and wet and has some issues, i'll see if i can make later a β version less messy ^^; (but there was a warning sign next to the centaur girl in the model card :D)