Marigold

Marigold repurposes a pretrained latent diffusion image generator (Stable Diffusion) for monocular depth estimation: instead of generating RGB, the denoising process generates affine-invariant depth maps, ensembling multiple stochastic passes for sharper results. Training needs only synthetic depth data, yet it generalizes zero-shot to real scenes — the reason the community noticed it. It was a CVPR 2024 oral and best-paper candidate.

The code is Apache 2.0 while the model weights carry a RAIL++-M behavioral-use license. A 2025 follow-up paper introduces Marigold as an affordable adaptation framework for other image-analysis tasks. It is slower per-frame than feed-forward depth nets like Depth Anything, but its fine-grained detail made it a favorite for 3D-lifting and depth-conditioned creative pipelines.

• Provider-reported (paper): zero-shot depth evals on NYUv2, KITTI, ETH3D, ScanNet and DIODE, competitive with or better than DPT and other zero-shot methods per authors
• No independent standardized benchmark leaderboard entry