mlboydaisuke/depth-anything-3-small-CoreAI-official

Depth Anything 3 (small) — Core AI export (official recipe)

fp32 export of depth-anything/da3-small via apple/coreai-models' official recipe (models/depth-anything/export.py, unmodified).

Runs out of the box with CoreAIKit's DepthEstimator:

let depth = try await DepthEstimator()   // downloads this repo
let map = try await depth.estimateDepth(for: cgImage)
let grayscale = map.cgImage()

Bundle layout

model/
└── da3-small_float32.aimodel

Graph contract

	name	shape	dtype
input	image	[1, 2, 3, 224, 224]	fp32, ImageNet-normalized (2 view slots — duplicate a single photo)
output	depth	[1, 2, 224, 224]	fp32, relative depth per view
output	depth_conf	[1, 2, 224, 224]	fp32
output	extrinsics / intrinsics	[1, 2, 3, 4] / [1, 2, 3, 3]	fp32

Preprocessing (224×224 resize + ImageNet mean/std + view replication) is handled by DepthEstimator.

Performance

M4 Max: ~7 ms per frame (GPU and Neural Engine are close). Requires macOS 27 beta / iOS 27 beta (device — the CoreAI framework is not in the iOS Simulator SDK).

License

Model weights: Apache-2.0 (ByteDance-Seed / Depth-Anything-3). Export recipe: BSD-3-Clause (apple/coreai-models).