Attempting to teach AI underwater stuff. Combines images from multiple sources including Amanchu!, Rou (rou22) and saver (artbysaver).

It is better at doing underwater stuff in general. This model can do underwater content in various styles and is not limited to the styles in the training data.

I recommend prompting for an artist style. Any style. The default style of the lora is a bit odd.

In negative prompt you can use "sepia" quite often. "Amanchu!", "monochrome", "greyscale", and "artist name" can help if they try to push through.

It can still do all the trained styles too. Saver, Amanchu! styles and Rou are better than the older individual loras.

V2:

Trained on Illustrious v1.0. Trained on 3207 images for 13400 steps. After repeats and batches this is 24 epochs of the primary trainig dataset. Images from other anime were added to stop model from learning "Amanchu!" style so strongly.

It is now better at generalizing style. Amanchu still often pushes through but not as badly as before.

Various things like diving masks, scuba gear, and diving poses are better.

Rou style is not broken anymore.

It is now a LoCon.

V1:

Training details: 942 images from Amanchu!, 634 images from Rou, 771 images from Saver.

Trained on Illustrious v0.1 for 5 epochs (5868 steps).