aarondevstack/DepthPro-512x512-coreml

DepthPro CoreML (512x512 Real-Time)

This repository contains the Fast (512x512) version of the DepthPro model, specifically optimized for the Apple Neural Engine (ANE).

DepthPro is a state-of-the-art monocular depth estimation model. This 512px version is designed for Real-Time Previews and high-speed video conversion on Apple Vision Pro and Apple Silicon Macs.

🚀 Key Features

• ANE Accelerated: Leveraging the Apple Neural Engine for ultra-low power and high-speed inference.
• Real-Time Performance: Ideal for interactive parameter tuning (Max Disparity, Convergence Plane).
• VisionOS Ready: Fully compatible with Apple Vision Pro via the DepthProPipeline.

📊 Performance & Requirements

Metric	Specification
Input Resolution	512 x 512 pixels
Compute Units	All (Optimized for ANE)
Average Latency	< 1.0s per frame (on M2/M3 chips)
Target Use Case	Real-time 3D Preview / Quick Video Conversion

This model is the best choice for the initial phase of your 3D conversion workflow, allowing for near-instant feedback while adjusting 3D rendering parameters.

📦 Repository Contents

The repository contains the following core components:

1. DepthPro_transform.mlpackage: Image preprocessing.
2. DepthPro_encoder.mlpackage: Feature extraction (ANE Optimized).
3. DepthPro_decoder.mlpackage: Multiresolution fusion.
4. DepthPro_depth.mlpackage: Final depth output.

🛠 Usage with Swift Transformers

You can download and cache this model dynamically using swift-transformers:

let hub = Hub()
let modelDir = try await hub.snapshot(repoId: "aarondevstack/DepthPro-512x512-coreml")
// Load models from the downloaded directory