Using ControlNet Depth to generate pngs with transparency?

[Susanoo1337]

Not sure how feasible it would be, but could a function be implemented that would use the depth map(or any other controlnet feature) to extract a foreground object and save a png without putting it back into a full image?

[cmeka]

Yes it's possible but sounds like it's outside the scope of this extension and even stable diffusion in general. Would have to be its own extension or gradio tool.

[brunogcar]

i would recommend https://github.com/thygate/stable-diffusion-webui-depthmap-script
if i understood you correctly you want to for example remove a character (foreground) from the scene (background)
it can invert remove background, Invert DepthMap (black=near, white=far),
you can select different models: midas and leres, etc
but if you want to do like photoshop, just remove and it would be replaced,you would have to inpaint what was removed

EDIT: i just read the title, transparency AFAIK isnt possible, currently there is no layer for that

[kft334]

Not sure if this is what you mean... This is easy to implement and works really well with hed/canny. Not so much with normals though.

Edit: My project is here in case you'd like to try it. It needs A1111 and sd-webui-controlnet installed to use. And a lot of the info on the itch.io page is outdated. The source code is included in the download: https://kfs334.itch.io/prompt-crafter-organizer

Edit: Works with blender too now. Only takes one click in blender to convert a scene to a ControlNet object with transparency.

[HolidayUser1]

I've been playing with settings and I am unable to graft in the fifth finger I need for the image. The image is clipart so I don't know how much that effects things. It always comes out 4 fingers...or a mangled mess.