facefusion-mobile-models (hyperswap 1b / 1c)

QNN context binaries for FaceFusion Mobile: the hyperswap 1b and 1c face-swappers, one binary per Hexagon tier. Float builds (no W8A16 quantisation): bigger files, exact output.

Files

file	tier	size
hyperswap_1b_v68.bin	v68	~388 MB
hyperswap_1b_v73.bin	v73	~388 MB
hyperswap_1b_v79.bin	v79	~388 MB
hyperswap_1b_v81.bin	v81	~388 MB
hyperswap_1c_v68.bin	v68	~388 MB
hyperswap_1c_v73.bin	v73	~388 MB
hyperswap_1c_v79.bin	v79	~388 MB
hyperswap_1c_v81.bin	v81	~388 MB

Tiers: v68 = Snapdragon 888 and older, v73 = 8 Gen 2 / 8 Gen 3, v79 = Snapdragon 8 Elite (SM8750), v81 = Snapdragon 8 Elite Gen 5 and newer.

Build

Same I/O as hyperswap 1a (target 1,3,256,256 + source 1,512, output 0), separate weights. Recipe: ONNX mask-drop + broadcast-Expand delete + fp16->fp32 demotion, then qnn-onnx-converter (float, no input list), qnn-model-lib-generator and qnn-context-binary-generator per tier (QAIRT 2.50, vtcm 8 MB, burst). 1b/1c prep verified at 90 dB against the raw graphs on held-out crops.

Use

In the app: Settings -> Models -> Import on the swapper row. The filename is the contract -- it must land under its own name in the app models dir.

Licence

hyperswap 1a is ResearchRAIL. The licence of the 1b/1c weights is not confirmed: check with the weights publisher before any commercial use.