Depth-Anything-V2 ggml

This is a single-file ggml implementation of Depth-Anything-V2.

Installation

ggml needs to be installed globally.

To compile the source code, use cmake:

mkdir build && cd build
cmake .. && make

Weights

Please refer to Hugging Face for the required model weights.

Model	Download Link
Small	Download
Base	Download
Large	Download

Usage

usage: ./dptv2 <s|b|l> <WEIGHTS> <INPUT_FILE> <OUTPUT_FILE>

Run DPTv2 model on an input image.

options:
  vit-size {s,b,l,g}    Specify the Vision Transformer size (`s`, `b`, `l`).
  weights WEIGHTS       Path to the input image.
  input INPUT_FILE      Path to the input image.
  output OUTPUT_FILE    Path to save the output image.

Example commands

• To save the output to a file:

  dptv2 s weights/vits.safetensors docs/input.jpg docs/vits.jpg
  dptv2 b weights/vitb.safetensors docs/input.jpg docs/vitb.jpg
  dptv2 l weights/vitl.safetensors docs/input.jpg docs/vitl.jpg

Example Images

Input

Output

---

Thanks to safetensors-cpp, stb and tinycolormap for their beautiful work.