I ported OmniAID image detection model to Apple's Neural Engine

Name: I ported OmniAID image detection model to Apple's Neural Engine
Availability: InStock
Author: woadwarrior01

by woadwarrior01·Apr 27, 2026·4 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidSlickBig Brain

Runs OmniAID detection locally on Neural Engine instead of uploading your data to servers.

Strengths

•Rewrote MoE architecture into static graph for Apple Neural Engine compatibility.
•Zero data leaves the device, solving privacy concerns of cloud-based detectors.
•Includes unicode cleanup tools to catch hidden AI text artifacts.

Weaknesses

•AI detection is an arms race; accuracy degrades as models improve.
•418 MB model size is heavy for a utility app.

Post Description

OmniAID is a hybrid MoE detector, so the PyTorch model dynamically routes each image through top-k semantic experts plus a fixed artifact expert. For the CoreML/ANE port, I rewrote that into a static graph. Every low-rank SVD expert path is materialized and the router turns into a dense gate vector where unselected experts have zero weights. That makes the graph much more ANE friendly while preserving the model’s behavior closely enough to ship a w8a16 quantized CoreML model (~418 MB). The quantized model gets to ~94.24% accuracy on the Mirage-Test dataset (also from the OmniAID authors).

https://arxiv.org/abs/2511.08423

https://huggingface.co/datasets/Yunncheng/Mirage-Test