neptun-org/wam-onnx

WAM MIT ONNX

ONNX Runtime exports of Meta's Watermark Anything MIT checkpoint.

These files are derived from the official MIT-licensed WAM SA-1B checkpoint:

• Source repo: https://github.com/facebookresearch/watermark-anything
• Source model card: https://huggingface.co/facebook/watermark-anything
• Source checkpoint: wam_mit.pth
• License: MIT

The non-commercial COCO checkpoint is not used.

Files

File	Purpose	Size	SHA-256
wam-mit-embedder.onnx	WAM embedder graph	4,608,338 bytes	DD5BD74F1E8C02FCE8D3BB1D964AEC3B314868A31FFDAD0DAEE1885B3DA61EB5
wam-mit-detector.onnx	WAM detector graph	373,825,798 bytes	022CD9D33B9F8AC1EDB1AB22F36299F5AAE5811B0303EA581079C85F16927CF0

Model Interface

The exported graphs use static 256x256 inference inputs.

Embedder

Inputs:

• image: [1, 3, 256, 256]
• message: [1, 32]

Output:

• watermark_delta: [1, 3, 256, 256]

Detector

Input:

• image: [1, 3, 256, 256]

Output:

• preds: [1, 33, 256, 256]

Export Notes

Local parity check against PyTorch:

• Embedder max absolute difference: 0.0001873
• Detector max absolute difference: 0.000016

These ONNX files are intended for native ONNX Runtime inference in applications that need WAM embedding and detection without a Python/PyTorch runtime.