Redid the component list and mixing methods to get more of that V3.0 feeling back, mostly via-vis face and squeezable cheeks. This version is more dynamic and uses perspective more than previous versions and will try to inject some sort of scene depth and dynamic composition into everything, sometimes to the detriment of the prompter if you're trying to go for a simple scene. You will rarely get head on staring into camera prompts, camera positions have to be clearly defined if you want just a simple 'looking at viewer, cowboy shot' generation. There is no scene composition backbone reinforcement like with V3.5D that was applied to make prompting similar to V3.2. The unets of this version, both conv and nonconv layers are different.
As SDXL is coming out soon, this is likely to be the penultimate version. I might release a new final version when Silicon30 comes out since Xynon confirmed that he finally revamped the codebase for autoMBW and he's about to start progress on it again, but otherwise everyone will probably move onto SDXL.
This has a baked VAE.
