Z-image support

[ifilipis]

Solved the scaling issue

• Native ComfyUI implementation limits the embedding scale to 1024x1024, resulting in distorted geometry on large resolutions.
• Z-image grid patches are now dynamically rescaled from base_res/output resolution ratio -> 1.0, creating a consistent image

Best verified by prompts with fixed geometric shapes (e.g. a grid), or high detail images, e.g. "New York birdseye view"

[lookuters22]

this fixes proportios but the image quality is bad. tested with all settings

[ifilipis]

this fixes proportios but the image quality is bad. tested with all settings

Try lower/higher dype_exponent - this is the only effective setting. I think, for portraits you want it lower than 2. Euler_ancestral also worked a bit better than res_multistep.

But yeah, this algorithm is not ideal for Z-image

[ifilipis]

This latest commit breaks vision_yarn, but improves on default yarn and ntk

Never mind, vision_yarn just works with res_multistep, and yarn and ntk work with euler