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