Thank you very much. To offer an explanation that might sound like an excuse:

1. I am currently using a 5060ti 16GB GPU.

Since this GPU lacks sufficient VRAM to train krea2,

I train the Transformer using NVFP4.

"Therefore, my LoRA lacks basic detail." + "I am a beginner and do not know superior training methods like others." = Consequently, "Honestly, my LoRA is terrible."

For this reason, I do not upload LoRA.

(I attempted several training sessions in Civitai as well, but they did not provide a satisfactory style.

My style involves conducting a considerable number of training sessions and making selections from among them.

In other words, it is a style that relies more on chance than precision.)

2. Why is the uploaded model NVFP4 instead of the more common 8-bit?

I am also aware that fp8 offers decent compression ratios and that int8_convrot provides excellent quality.

However, due to the reason mentioned in point 1, the better the Base model's quality, the more significantly the shortcomings of my low-quality LoRA become apparent (I have tested this myself).

In my experience, the NVFP4 compression model in the krea2 environment has the characteristic of 'omitting some details' when representing images compared to the 8-bit compression model.

Since this characteristic helps to dilute the first drawback of LoRA, I uploaded only NVFP4 to hide various disadvantages.

Thank you