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
First release.
FAQ
Comments (16)
What is the difference between convrotINT8 and regular INT8?
Int8-ConvRot is supposed to have higher accuracy when quantized down; so less visual degradation when optimizing for lower-end hardware.
Supposedly its higher accuracy than INT8 and FP8; comparable to FP16 and BF16.
On my 5060TI 16GB Int8 is slightly faster than Int8ConvRot, which is pretty much the only thing I care about.
I provided all those formats because of different hardware. Didn't provide GGUFs yet because I had issues with converting to Q8 and Q4 formats.
Technical speaking, standard 8-bit compression (plain INT8) usually destroys image quality and may cause blank "black images" because rounding decimals to whole numbers creates mathematical errors.
The ConvRot matrix rotation fixes this. Before the numbers are rounded, a special mathematical rotation (a block-Hadamard transform) is applied to the layers. This mathematically spreads the errors evenly across the network, ensuring that the visual sharpness and colors of the original model are preserved.
Good day! I came across your model, downloaded it, and I'm currently in the testing phase. Based on my first few prompts, the model is delivering great results. I hope you continue to develop it. One major plus I've noticed right away is the skin texture and realism—straight out of the box, without any LoRAs, the model produces excellent results in my opinion. Thank you for sharing it! I'll share the results of my tests later.
Thank you, I do plan on keeping it updated.
Кстати да. Krea 2 Identity Edit с помощью этой модели на ура перенесла все прыщики и дефекты кожи. Позже, я создал датасет и лору. Текстуры сохранились.
Hey, this is an Alpha? It works very well, for my Opinion.
I´m use 5+2 Steps.
1.Pass = Euler ->Beta57
2.Pass = Euler Ancestral -> Beta57
No Lora used.
I think i get good Results.
Thanks for Sharing Bro.
I wouldn't even call it an Alpha; I experimented with the model for a few days and this is what came out of those experiments. I do plan on fine-tuning it further, still working on a dataset at the moment.
@Part_LoRAs I'm excited to see the future result. Great Work :)
Модель очень даже неплоха. Но что то мне подсказывает, что вместо версии int8convrot - обычная fp8. Вы конвертировали с помощью Starnodes Model Converter?
ps. Переконвертировал вашу основную модель в int8convrot с помощью вышеуказанной ноды. И действительно, генерация залетала практически в два! раза быстрее. Без потери качества. И это учитывая мои 8gb VRAM.
If that's the case, I hope the author fixes it.
I converted all models using silveroxides convert_to_quant tool. If the ConvRot .safetensors is broken in any way then its convert_to_quant fault. I don't have time right now to check but I'll take a look soon.
Its fixed, Int8 ConvRot is correct now. Not sure what exactly happened, maybe I renamed the files wrong. I made sure to inspect the model to confirm its in the correct format.
The file is being uploaded, it will be available soon.
@Part_LoRAs Thank you
Surprisingly wouldn't let me use with Int8 diffusion loader. Worked with regular though, and it is faster than the other official INT8 I have, so I suppose that's a huge win.
Additionally this is the only custom INT8 I've found to work so far.
Maybe it doesn't recognize that its ConvRot? Who knows.