It looks like it could be combined with Patchfusion, according to their paper, is a high-definition approach to depth maps that is technically independent of the base model.

Great; Marigold produces depth maps of the same revolution as input

Wow, Marigold inference is using a diffusion pipeline to produce depth maps? I think this is heavy enough that we make it an optional component. (Or we want a component management system for ControlNet?)

Yes, but it can be integrated similarly to https://github.com/kijai/ComfyUI-Marigold by copying the code, so there is no need for extra dependencies. When we roll out an update for Marigold, we will notify dependency repositories about improvements upstream!

Yes, but it can be integrated similarly to https://github.com/kijai/ComfyUI-Marigold by copying the code, so there is no need for extra dependencies. When we roll out an update for Marigold, we will notify dependency repositories about improvements upstream!

Apologies for the late reply, because while huggingface's range of packages are very good to replicate and implement, the iteration between versions makes breaking changes difficult to maintain.
For example diffusers 0.20 and 0.21 are not common between them, nor are 0.24 and 0.25.
We try not to use diffusers in webui, and we're trying to figure out how to read the diffusers config file to replicate the neural network constructs.

Wow, Marigold inference is using a diffusion pipeline to produce depth maps? I think this is heavy enough that we make it an optional component. (Or we want a component management system for ControlNet?)

I think a generic config neural network structure for reading diffusers might be the way to go.

Supported in https://github.com/lllyasviel/stable-diffusion-webui-forge