RIFE v4.6 for the Hexagon NPU (HexAnima)
Practical-RIFE v4.6 (hzwer, MIT) video frame interpolation converted to QNN HTP context binaries (Hexagon V69, Snapdragon 8 Gen 1 / 8+ Gen 1, fp16) for the HexAnima Android app (https://github.com/ayanoby/hexanima): Upscale -> Video, x2 / x4 frames per second.
One file per frame size, rife_t<H>x<W>.bin: input [1, 6, H, W] (two RGB frames in 0..1, channels 0-2 and 3-5), output [1, 3, H, W] (the frame between them). The sides are multiples of 32; the app pads a video with its edge pixels to the smallest file that covers it. The time is about 0.04 s plus 0.45 microseconds per pixel for one new frame.
manifest.json lists the size and sha256 of every file; the app downloads only the size a video needs. Checked against the original fp32 model on the phone: 67.6 dB (256 x 448 test pair, practically identical). These files only run on that chip generation; they are not general ONNX/PyTorch weights.
Converted from flownet.pkl (v4.6, IFNet_HDv3) with the warp written as grid_sample, exported to ONNX (opset 17), qairt-converter (batch 1, fp16) and qnn-context-binary-generator. License: MIT (see LICENSE, Copyright (c) 2021 hzwer); these are modified (converted) copies of the original weights.