Marigold is a free, open-source computer-vision model from ETH Zurich that estimates detailed depth maps from a single photograph. A CVPR 2024 oral paper and best-paper-award candidate, it repurposes the image-generation power of Stable Diffusion for monocular depth estimation, achieving state-of-the-art zero-shot accuracy on scenes the model has never seen.
Marigold is an open-source project on GitHub with about 3,200 stars, actively maintained into 2026. Instead of training a depth network from scratch, Marigold fine-tunes a pre-trained latent diffusion model on synthetic data only โ yet generalises remarkably well to real photos “in the wild”, outperforming dedicated depth estimators on standard benchmarks in zero-shot settings. The project has expanded beyond depth: the Marigold family now includes variants for surface-normal prediction and for intrinsic decomposition (separating an image into albedo and shading). The repository provides pre-trained checkpoints, a simple inference pipeline and a Gradio demo, so you can get depth maps from your own images in minutes.
Marigold is completely free and open source under the Apache-2.0 license. Download the checkpoints and run it on your own GPU at no cost.
Because it inherits Stable Diffusion’s image priors, Marigold handles fine structures, reflections and unusual scenes more gracefully than many purely discriminative depth networks, and its ensemble inference option trades a little speed for noticeably cleaner, more detailed depth maps.
Marigold is for computer-vision researchers, 3D and AR developers, and photographers’ tooling builders who need reliable depth from a single image โ for scene reconstruction, relighting, effects or robotics perception.