This is supposed to be a LORA used for generating Kemono content. The original purpose of this model is to experiment with new algorithms. Please respect the artists.

Usage

Model : This model was trained on NoobAI 1.1. It also works with V-pred 0.65S. You can choose any model that derived from Noob 1.0 or 1.1 I guess.

Tags : There are some samples above. For some reasons, I don't want to publicly display all the names of the artists here. You can download and check the metadata to get the list of artists who have been trained.

You can try using the built-in artist tags in NoobAI to improve the quality.

Limitations

Not using artist tags may lead to a decline in quality.

The background is usually blurred. I think this is an issue with NoobAI.

Weird pov hands coming out from everywhere and touching my kemono bois; if you're concerned about this, please add "pov hands" and "pov" to the negative prompts.

Bad at clothing details.

Training details

LyCORIS type : LoCon (Linear Rank 32, Alpha 1; Conv Rank 8 (for R1) 16, Alpha 1)

Optimizer type : ProdigyPlusScheduleFree (prodigy_steps = 25% step, weight_decay = 0.1, use_bias_correction = True)

LR : Both 1

LORA+ Ratio : Both 16

Mixed Precision : BF16

LR Scheduler : Constant

Dropout (Network/Rank/Module) : 0.1

Trained with FP8 base, with a batch size of 2, VRAM does not exceed 8GB. It can be useful if you are using a GPU with 8GB VRAM.

The optimizer and fp8 base settings are crucial, it can significantly reduce VRAM usage compared to original Prodigy.