Seen this on Twitter, very cool! > We present Marigold, a diffusion model and associated fine-tuning protocol for monocular depth estimation. Its core principl

Seen this on Twitter, very cool! > We present Marigold, a diffusion model and associated fine-tuning protocol for monocular depth estimation. Its core principle is to *leverage the rich visual knowledge stored in modern generative image models*. Our model, derived from Stable Diffusion and fine-tuned with synthetic data, can zero-shot transfer to unseen data, offering state-of-the-art monocular depth estimation results.