Video-Depth-Anything-Large-MLX

Video Depth Anything Large (MLX)

MLX port of Video Depth Anything (ByteDance, CVPR 2025 highlight): consistent monocular depth estimation for arbitrarily long videos. Converted from the official checkpoint depth-anything/Video-Depth-Anything-Large.

Architecture: DINOv2 backbone + DPT head with AnimateDiff-style temporal motion modules. Outputs per-frame depth maps, not text.

Usage

from mlx_vlm import load
from mlx_vlm.models.video_depth_anything.generate import (
    VideoDepthPredictor,
    read_video_frames,
)

model, processor = load("mlx-community/Video-Depth-Anything-Large-MLX")
predictor = VideoDepthPredictor(model, processor)

frames, fps = read_video_frames("input.mp4", max_len=300, target_fps=15)
depths = predictor.infer(frames)  # (T, H, W) float32

Notes

• Inputs are channel-last (B, T, H, W, 3); H and W must be multiples of 14.
• Metric model: no. Metric models output absolute depth (meters) and skip scale/shift window alignment.
• Weights are fp32. On GPU, output matches the PyTorch reference to ~1% relative (Metal fast-math); on CPU to ~1e-5.

License

CC-BY-NC-4.0 (same as the source checkpoint). Non-commercial use only.