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

The result is close to base model, helps with detail and stabilization.

The recipe is different from the previous test. IterComp is added at 2.0 value, then a slerp is done with the base model at 0.5. Then, an orthogonal procrustes with the result and the base model, to keep the original composition.

[Click for information on the model & how to use it].

-- usage tips --

I recommend using RescaleCFG at 0.7 to prevent oversaturation. For upscaling, this latent upscaler preserves the latent quite well, and Manhwa Null removes jpeg artifacts.