Marigold-LCM: fast monocular depth estimation based on Stable Diffusion

Follow the full discussion on Reddit.
Marigold is a universal monocular depth estimator, delivering incredibly sharp predictions in the wild. Based on Stable Diffusion, it is trained with synthetic depth data only and excels in zero-shot adaptation to real-world imagery. The original method required tens of seconds to process the original image, which was justified given the boost in depth map quality. Following in the footsteps of text-to-image research, we performed latent consistency distillation of the original model, giving an order of magnitude speed-up. Check it out!