Recently, I was playing with a workflow that required passing generated images through a depth estimation model and I thought: why not generate depth maps directly?

The results have been, well, favorable. It seems to struggle with some objects that are black/white naturally, and it also tends to give it too much detail sometimes. With that said, it might be still useful.

My use for this is as follows:

To start, I generate a few images around what I want, ending up with something like this:

Which I can then bash around and edit on Photoshop to get something I like. Since it's all in grayscale, that makes it much simpler to work with. After that, I'll end with something like this:

With that, I can do another T2I run with a depth ControlNet, and that will give me a result like this:

The point is primarily to simplify the editing work, and to avoid the artifacts from my usual I2I workflow.

Hopefully that can be useful to some of you as well!