This is exactly how model releases should be documented. I wish everyone did it the way you did

I am gonna be honest , I did not try your model yet ,, but since you said you are going to release the V2 today , i will wait for it and try it ,, and give you my honest opinion .

Thank you for sharing your work with us . Have a nice day

The only downside to this amazing model is that it results in a Western-looking face due to a lack of training for Asians 😂🤣🤣

looks great.. thanks for all your talents. * I think the names/files of both formats you uploaded (.safesensors) -INT8 and FP8 formats above are the same. They are at least the same name.

from where to download the abliterated text encoder which you're using with this model

For people with the wet / sweat / cum problem this is NOT a problem of this checkpoint, is a Krea 2 problem, I tested most ckpts and LoRA's, looks like it is a internal issue that is tied to LoRA's and all custom checkpoints.

does this new version work with other loras?

This Model (INT8) is surprisingly fast compared to some of my other INT8 Models. Well done!

Good evening, does anyone have the workflow?

The workflow in the description isn't working for me; I'm new to this.

Does v2 holds better character loraas?, v1 didnt.

For me v2 is less realistic than v1, v2 produces 3d looking outputs.

In my opinion v1 is better than v2, v1 is real life as fuck and i am getting plastic skin with v2

V2 is not that bad, you can get nice outputs without an elaborated prompt. While V1 is like uncensored krea2. Maybe for V3 you can merge both to see what happens. :]
Anyway, thanks for sharing. keep it up. ♥

Can you reupload your int8 convrot quant for v2 pls, it has txtfusion layers converted in fp8 which make this quant not working in ComfyUI-INT8-Fast custom node (I have to use it because of triton)
Also I suggest not to quantize txtfusion layers at all, like redcraft model maker do

Alrighty, will give my opnion (keyword mine), this is now by far my favorite model, the prompt adherence is there which is what's important, the quality is also there as the images come sharp by default on 1536p or 2048p.

And it understands the keywords asian, japanese, korean and chinese,, which strangely some Krea 2 models don't support, it also supports gingers and dreadlocks and on top of all that I stacked over 6 LoRA's on top of it and it played nicely with then :)

any chance you can upload an nvfp4 or an fp16 so we can convert ourselves? Some of us struggle with running 12gb files.

Are there any plans to release a raw version so I can use Turbo Lora myself? In my experience, raw + turbo often produces better results than turbo checkpoint. :)