LyCORIS models trained on 20 images from @shinken_gomi on Twitter (now X).

License: Fair AI Public License 1.0-SD

Rambling

The XL version is my first successful attempt at training an SDXL LoRA with a reasonable amount of time, thanks to FP8 base training and Prodigy optimizer. The SD1.5 version is an improved version of the initial training runs for this specific dataset.

The dataset for both models is originally intended to be used as a testing ground because of its level of detail and diversity. There are better datasets available, but I actually quite like this one.

As it is a style model, you are encouraged to use it in conjunction with other models rather than using it alone.

Also, I'm slowly transitioning away from SD1.5. The main reason is that I want my maid catgirls to look cute and images of them to be high-resolution.

Model specs

SDXL

Base model: Animagine XL 2.0

Recommended weights: 0.8-1.1

Train Text Encoder: Yes

Training steps (equivalent): 400 (800)

Rank - Alpha - Conv - Conv Alpha: 16 - 32 - 8 - 16

Training resolution: 768x768 (bucketing)

SD1.5

Base model: AnyLoRA

Recommended weights: 0.5-1.0

Training steps (equivalent): 370 (1480)

Rank - Alpha - Conv - Conv Alpha: 64 - 32 - 32 - 16

Training resolution: 640x640 (bucketing)

Dataset specs

Tagging: Automatic (WD14-moat-v2)

Filtering: Manual + Automatic (Cafe Aesthetic)

Crawler: JDownloader 2

Cutoff: December 2023