Samples generated using nvfp4 DIT and fp8 TE.
The main objective of this model is to have a base uncensored model for further fine-tuning that is still very close to the base Krea 2 Raw/Turbo.
I'm working on a larger dataset at the moment. No ETA on when it will be trained. Feel free to make suggestions for things to train.
Sampler & Scheduler: ER_SDE + Beta or Euler + Simple
Steps: 8-12
CFG: always at 1.0
Shift: the Krea 2 official repo uses 1.15 for Turbo.
Description
Attempt to fix some anatomy, textures and prompt adherence issues.
All quantized .safetensors files converted using SVD and convert_to_quant tool.
FAQ
Comments (7)
hoooly shit man, how did you train this model so well? the v2 gives me photorealistic results not a single lora combination or checkpoints offers
please go on 😳💪
Thanks for the kind words.
I'm currently testing V2 BF16 and i do like it. well trained model, i also use it with ComfyUI-Krea2T-Enhancer and Krea 2 Conditioning Control, i do get great results for both Sfw / Nsfw. good job :).
Thank you.
Hello ! just a little feedback, i like the quality of your model, its ability to render dark skin tones and such, there is a few issues i'm seeing though, maybe flaws from krea2 itself or from my prompting, or from the training dataset :
- cannot get a deep penetration whatsoever, it's just the tip even with deepthroat lora
- the "pov" is overfit on "pov kneeling leaning back", it's insanely hard to get any other pov than this one (lying on back is ok)
- it's somewhat overfit on doggystyle, producing it more likely even on totally unrelated poses, every time we mention "back" or "behind".
A list of concepts i haven't been able to make work with pov + anal sex basis :
- from behind/all fours/doggystyle (no anal 99% of the time)
- prone bone seems to be unknown (goes doggystyle instead)
- facing (pov missionary or pov mating press, or pov anvil [...etc], anal or not, i cannot get them)
- amazon position
And before you ask, no i'm not using tags, i'm using natural language, it's just shorter here to explain you the idea with tags than with complete sentences.
Weirdly, the piledriver + pov works flawlessly, while no other previous model i tested was able to do it.
Thanks for sharing your work, hope to see more of it !
I did not train any specific positions yet so it makes sense it can't really do it.
I'm working on a dataset to fix most of those issues but I cannot promise a timeline at all. Its just been really slow so I don't just burn out as always.
Whatever you are able to generate right now is VERY likely to be something that Krea 2 base model already understands. The goal here was a base uncensored model (that is really close to base Turbo) that I can use for further tuning.
For now, you are expected to use LoRAs for more specific things.
My dataset folder is currently divided into 40 categories which aren't even captioned yet, I'm still gathering images and clearing their watermarks.
After some testing, you can get most of those by just prompting.
The anal issue is one I already knew about and should be fixed in a future update. It was expected because Krea 2 100% was trained with NSFW in the pre-training phase but it also saw generic descriptions for all of those images, and vaginal is heavily over-represented; I don't think it ever saw a caption that had "amazon sex position" or "prone bone" or "anal penetration" at all... they definitively trained it on NSFW content tho and my model is supposed to be as close to the base model as possible but uncensored. That also means you are better of with longer detailed prompts.
The Amazon position can be prompted by being highly specific and detailed about the position, but the Turbo model almost always "self-correct" to an easier path mid-way through a generation; the preview clearly shows the general composition for the concept but then it self corrects by placing the female character on top in a straddling position and just leaves the guy legs up in the air. I did however get it quite easily by using an unreleased version of Neutrino that is based on the RAW model instead of the Turbo. I was using a Turbo LoRA at 0.6 weight + Neutrino RAW.
A suggestion I have (and its something I use myself) is to use a LLM to caption a reference image for whatever it is that you are trying to generate to see exactly the kind of description that was likely used in Krea training.
I'd be interested in seeing the prompts you used if possible.