Video-Depth-Anything

Why we included this project

Per-frame depth from single-image models tends to flicker across video, which breaks downstream tasks like 3D reconstruction or VFX. Video Depth Anything keeps depth stable over arbitrarily long clips by carrying temporal information between frames, and it adds an experimental streaming mode for continuous input. Pretrained small, base, and large checkpoints, plus metric-depth versions, run through a single script that outputs grayscale, NPZ, or EXR files. It's lighter and faster than diffusion-based video depth methods, so long footage stays practical on one GPU. The repo also bundles inference, evaluation, and training-loss code, making it a solid starting point for teams building scene-understanding tools.