Starting with one of my private realism models this was trained on sprites extracted from PS1, Saturn, and N64 games to serve as the foundation for a new branch of pixel art models that should be ideal for game development.

The model was trained on a wide range of common fantasy, sci-fi, and numerous proprietary concepts to sub 0.07 and aggregate sub 0.03 losses from source, I then readded ~40% of the starting realism model to improve quality and continuity for larger resolutions without sacrificing the desired behaviors.

The plan from here is to train on various datasets to make specialized versions of itself that do niche things.