Video Depth Anything Small (MLX)

MLX conversion of depth-anything/Video-Depth-Anything-Small (ByteDance, CVPR 2025 highlight). Consistent monocular video depth estimation: DINOv2-vits backbone + temporal DPT head.

Converted with python -m mlx_vlm.models.video_depth_anything.convert --variant small.

Usage

from pathlib import Path
from mlx_vlm.utils import load_model, get_model_path
from mlx_vlm.models.video_depth_anything.processing_video_depth_anything import (
    VideoDepthProcessor,
)
from mlx_vlm.models.video_depth_anything.generate import (
    VideoDepthPredictor,
    read_video_frames,
)

path = get_model_path("jarjoura/video-depth-anything-small-mlx")
model = load_model(path)
processor = VideoDepthProcessor.from_pretrained(path)
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 depth maps

Validated against the PyTorch reference: ~1e-5 max relative error on CPU, ~1% on GPU (Metal fast-math).