Marigold Depth v1-1

Marigold-Depth-v1-1 is a generative latent diffusion-based model for affine-invariant monocular depth estimation from a single image. It is fine-tuned from Stable Diffusion 2 and retains its strong visual priors, enabling state-of-the-art zero-shot generalization across diverse domains without additional training data.

Technical Details

• Output: An affine-invariant depth map with values between 0 and 1, representing relative near-to-far ordering. When ensembling multiple predictions (ensemble size larger than 2), an uncertainty map is also produced.
• Resolution: Designed for an effective resolution of approximately 768 pixels on the longer side. For optimal results, larger inputs should be resized accordingly.
• Scheduler and Steps: Works with the DDIM scheduler using between 1 and 50 denoising steps.
• Training: Fine-tuned exclusively on synthetic data over a few days on a single GPU, as described in the CVPR 2024 paper and the journal extension by Bingxin Ke, Kevin Qu, Tianfu Wang, Nando Metzger, Shengyu Huang, Bo Li, Anton Obukhov, and Konrad Schindler.

Performance

Marigold achieves state-of-the-art zero-shot performance across a wide range of benchmark datasets. In specific cases, it delivers over 20% improvement over prior methods. This model is hosted by gigarouter as a managed, OpenAI-compatible API for easy integration.