Gave up on ICON/PaMir/PiFu/Eva3D ever working on my computer and tried to make it work on my own.
Workflow: Stable diffusion, Automatic1111.
-Text2Img something that your model know the back and front of with a prompt.
-Take image, feed into Depth map script, output Left/Right stereogram
-Cut Stereogram into two pieces, feed either left or right into img2img with same prompt, low but non-zero de-noising, and depth map script again. Repeat on same side.
Still need to play with the settings (de-noising and depth shift), up the resolution, change the prompt a bit while rotating. Hope someone else has better luck than I did. Anyone have a better workflow?