Add DA3MONO-LARGE Core ML (504x504, FP16)

3961438 verified 4 days ago

2.06 kB

license: apache-2.0
base_model:
  - depth-anything/DA3MONO-LARGE
base_model_relation: quantized
pipeline_tag: depth-estimation
library_name: coreml
tags:
  - coreml
  - depth-anything-3
  - monocular-depth
  - macos
  - apple-silicon
  - fp16

Depth Anything 3 Mono Large — Core ML (504×504, FP16)

Core ML conversion of DA3MONO-LARGE (Depth Anything 3, ByteDance Seed) for on-device monocular depth on Apple Silicon Macs. It runs on the GPU or the Neural Engine.

Anyone may use it under the same Apache-2.0 license as the upstream weights.

Files

File	Size	Notes
`DA3MonoLarge504.mlpackage.zip`	~620 MB	ML Program, FP16 weights, minimum macOS 13

Specification


Input `image`	`(1, 3, 504, 504)` float32, RGB in `[0, 1]`. ImageNet normalization is built into the model.
Output `depth`	`(1, 504, 504)` relative depth (larger = farther)
Output `sky`	`(1, 504, 504)` sky score (values above 0.5 are sky)
Shape	Fixed at 504×504. For other aspect ratios, letterbox (pad) to a square.

Speed on an M5 Max is about 31 ms per frame on the GPU, measured end to end including I/O.

On a sample frame, the output differs from the PyTorch reference by about 0.03% (mean relative difference).

Conversion notes

Sky post-processing: The upstream model's data-dependent sky step is not part of the graph. Use the sky output to set those pixels to the farthest depth.
Position embedding: The DINOv2 position embedding was interpolated to 504×504 and stored as a constant before conversion, because Core ML has no bicubic resize.

License

Released under the Apache License 2.0, the same license as the upstream DA3MONO-LARGE weights (see LICENSE).

This is an independent conversion. The upstream Depth Anything 3 authors are not affiliated with it and have not endorsed it.

Credits

Depth Anything 3 by ByteDance Seed (Haotong Lin et al.).