Depth Anything 3 Small Multi-View Core ML
Copyright 2026 Shuichi Tsutsumi

This repository contains Core ML conversions of the Depth Anything 3 Small model weights.

Original work:
  Depth Anything 3 Small (depth-anything/DA3-SMALL on Hugging Face,
  revision e08cab65ca0ec38e7826075418411ab90cab4da3), licensed under the Apache License, Version 2.0.
  The original repository has no NOTICE file.

Changes made in this conversion:
  - Converted from PyTorch to Core ML packages (.mlpackage) with coremltools.
  - Multi-view input with camera poses: the models take 2 or 3 images together with each image's
    extrinsics and intrinsics (for example from ARKit), instead of a single image.
  - Fixed input sizes of 504 x 504 or 392 x 392 pixels.
  - In the "sliced" variants, the global attention across views is computed in slices of the query
    (16 slices, or 8 in "sliced8") to keep the intermediate matrices small.
  - In the "headcombo" variants, the depth head was rewritten (transposed-convolution upsampling, and the
    upsampling moved later in the head). Depth values differ slightly from the original at object edges.

The full text of the Apache License, Version 2.0 is in the LICENSE file.