Inclined Pixels

Pretty simple dreambooth meant to make generations into an inclined pixel art. Pixels are not perfect, but it works quite well.

Loras are working with this model, and are working quite well, even some of my heaviest ones are transferring very nicely.

Issue of edge pixels exist(most prominent on top edge), but, eh. I have an idea of how to set up a process to fix that in dataset, so, maybe in future versions, if model will be interesting to people.

NSFW works, but it was not trained to do anything in that regard.

Training info i guess

Simple 831 image dreambooth with tuned parameters.
Decoupled Learning Rate, decoupled Weight Decay, tenc clipping, etc.
Tenc stopped at 75%.
Initial offset noise is 0, noise travel is 0.0314.
Model config was unfrozen, token limit was set to 150, but dataset is manually tagged, and i don't recall such long strings there.
100 epochs, 83100 steps. Possibly could be tuned at higher res for 10-20 more epochs.

Dataset used belongs to me and was not scraped from the internet.