melmass/liveportrait-decoder-coreml

LivePortrait SPADE decoder, Core ML

LivePortrait's SPADE decoder (KwaiVGI) as a Core ML ML Program in float16, built with coremltools' MIL builder from spade_generator.pth

• Input feature: float32 [1, 256, 64, 64], the warping network's output (NCHW).
• Output face: [1, 3, 512, 512] in [0, 1].
• Instance norms read their input 256× smaller and the res-block convs 64× smaller (results scaled back): the activations reach ~10⁴, past float16 otherwise.
• Run it on the GPU (.cpuAndGPU): 32 ms on an M5, 0.15–0.21 levels mean from the float32 decoder. The Neural Engine is slower (41 ms) and 3 levels off.