MiDaS

Before Depth Anything and Depth Pro, MiDaS from Intel's Intelligent Systems Lab was the go-to model for estimating depth from a single photo. Its central idea, training on a blend of up to a dozen datasets with a loss that ignores scale and shift, gave it unusually robust zero-shot results, and its DPT transformer variants became a stock component in ControlNet depth conditioning, 3D photo effects and research code. Development effectively ended with MiDaS 3.1, whose weights were published in December 2022; the only later change was a security policy file in August 2024, and Intel has since archived the repository. Everything still works: release checkpoints download from GitHub, the torch.hub entry loads them, DPT versions live in Hugging Face Transformers, the small models run on CPU, and the MIT license permits commercial use. Keep in mind it outputs relative inverse depth, not distances in meters. For new projects, Depth Anything V2 or Depth Anything 3 (small models Apache-2.0, larger ones non-commercial) give sharper edges, and Apple's Depth Pro or the metric Depth Anything 3 model provide true metric depth.