Stable Diffusion 2.0 introduces depth2img, which is like img2img but uses depth map estimates to help guide the generation. This should help in situations where img2img doesn’t understand how to segment foreground and background, creating weird artifacts.
@Riedl very cool that this is self supervised