Using sd_mecha (recipe: part 1, part 2).

This merge attempts to fix stability issues that occur in 0.75 because of the text encoder not being in sync with the unet during training.

The unet and clip of 0.5 is added with the rotate method, which is similar to orthogonal procrustes approximation. I also find that this merge has better saturation than either.

Recommended method for high-res:

Scaling: 1.5x (optional latent upscaler)

Model: 4x-NomosUni-span-multijpg

0.5 strength with this yolov8 model