Krea Noct
A photoreal Krea 2 merge for NSFW work. Explicit prompts work with no bypass LoRA.
Prompt examples
The prompts of three showcase images, written like captions of the finished picture: the medium first, then the woman, what she does, where, the light and the framing. All three use 8 steps, cfg 1, euler, simple, 1024x1536.
Treehouse, seed 97032
A phone photograph of a nude woman in her twenties with strawberry-blonde hair in a loose braid and an athletic build, lying on her stomach on a patchwork quilt on the plank floor of a wooden treehouse, her chin on her folded hands, her bare feet raised and crossed, looking out of a small square window at green leaves. Dappled sunlight through the window across her back and legs, a wicker basket and a string of unlit fairy lights along the wall. Full body from the side at floor level, natural skin texture.
Sea rocks, seed 97112
An amateur photograph of a nude woman in her twenties with wet wavy brown hair and an athletic build, lying on her stomach on a smooth flat granite rock by the sea, propped on her forearms, looking out at the waves breaking below. Bright midday sun, turquoise water and white foam around the rocks, a striped towel and sunglasses beside her. Full body from a slightly low angle along the rock, natural skin texture.
Anime cafe, seed 97211
An anime illustration with clean line art and soft cel shading of a woman in her twenties with long auburn hair and a cream knit sweater, sitting at a window table in a cozy cafe on a rainy evening, holding a cup of cocoa in both hands and watching the rain, a grey cat asleep on the table beside her. Warm cafe light inside, blue rain streaks on the window, a slice of cake on a plate. Medium shot, thin dark outlines on her hair, face and sweater.
What it does well
Full-body and seated nudes with natural bodies
Phone-snapshot and flash looks, daylight interiors, outdoor scenes
Anime, illustration and 3D too, no LoRA (see the gallery)
Plain-caption prompts: who, what she does, where, the light
Which file
12 GB and up: int8 convrot
8 GB cards: int4 mixed, about 1.4x faster
fp8 and bf16 for merging and training
Settings
8 steps, cfg 1, euler, simple, 1024x1536, no upscale needed. Text encoder qwen3vl_4b_fp8_scaled, VAE qwen_image_vae.
Notes
8 steps is the sweet spot; 12 or 16 steps sharpen hands and small text
Describe the face in one clause and it stays consistent across seeds
Full-body framing and an off-camera gaze give the most natural results
中文：Krea 2 写实无审查融合模型，无需绕过 LoRA 即可生成裸体与性爱场景，也能画动漫与插画。8 步、cfg 1、euler/simple、1024×1536。int8 convrot 需 12G 显存，int4 适合 8G。
Description
V2: re-merged on ULTRA with the Kroma text-fusion core kept intact. Same settings as V1 (8 steps, cfg 1, euler, simple). int8 convrot, int4 mixed, fp8 and bf16.
FAQ
Comments (13)
Could you add prompts to your examples. I am so bad at this, it's not funny anymore :D
Sure, here's one in the style it likes (a plain description: shot type, who, what she does, where, the light): "A candid photograph of a nude woman in her twenties with a long black braid, sitting on a stone step at a quiet beach at sunset, hugging her knees and looking out at the waves. Warm low sun from the right, wet sand and a towel beside her. Full body from the side, natural skin texture." The settings are in the page description.
Just way to many merges to try... This one looks OK, but how does it handle character LoRA and style/illustration LoRA?
You just have to test them unfortunately otherwise you'll never know. I have an automated setup that cycles through dozens of models for these tests. I can tell you the models fall into the following categories from extensive testing:
- truly unique ones usually half censored like Krealism or Pure Tukano
- top of the line, like Fineporn - this one here is also some kind of merge of top of the line models so it's very high quality
- middling models that try but don't reach the same quality
- simply botched models that might be impressive if you don't know any better like Redcraft
That said depending on what you prompt the output is vastly different so even to test them you first have to find prompts that work, since most prompts just fail as a prompt and that's Krea not any particular model. Have fun.
To give you an example I'm running a 33 prompt test on 39 models right now with 10 images given for each prompt. I have no idea what the results will be until I check back in a few hours.
@ferrrett33 wow, comprehensive! I will give it a test. It’s just challenging with my low VRAM, so testing takes a lot of time.
@xFennec777 yes it does take a lot of time. I have 8GB VRAM and I've been working on it for months, automating the whole thing in Comfy so 95% of it is no longer manual, I just have to guide the process like 5% of the time.
It takes Krea 2 LoRAs, but how well it works depends on the LoRA, so it's case by case.
Great model. one of the ones that best follows the prompt and accpet Loras
Glad it works well with your LoRAs!
It's a good model overall but I find v1 better than v2 which produces a lot more anatomical monstrosities when it comes to NSFW content. Also it suffers from the same problem as almost all KR2 models: it prioritizes anal penetration over vaginal penetration and forces the appearance of bodily fluids in unprompted for areas, even if you've prompted for a specific location.
I ve posted a set and all the cum generated on the character's crotch or buttocks is forced by the model. Furthermore, i can only achieve proper vaginal penetration, both anatomically and in terms of positioning, with a dedicated LoRa like SNOFS. Otherwise, the result is rarely satisfactory.
Last but not least, text-fusion isn't a consistent solution to these issues. So as said above you ll need to use a dedicated LoRA for proper vaginal penetrations.
对我来说这是优点
@kkk1415 It's a matter of preference. Anyway, a good model isn't one who favors anal over vaginal, or vaginal over anal. A good model is one who can do what you prompt for, and nothing else.
Thanks for the feedback. Good to see in your newer comment that V3 does better.