IFRNet-S for the Snapdragon NPU (Nightmare Mobile)

QNN context binaries of IFRNet (Kong et al., CVPR 2022, MIT), the small model with its GoPro checkpoint, for the Smooth motion switch on the Wan video nodes of Nightmare Mobile: one new frame halfway between every pair, so a 16 fps clip plays at 32.

  • ifrnet_s_<W>x<H>_g16_<arch>.bin: sizes 832x480, 480x832, 1280x736, 736x1280; v79 (Snapdragon 8 Elite, also run by newer chips) and v75 (8 Gen 3). QAIRT 2.49.
  • Inputs img0, img1, output imgt: [1,3,H,W], RGB in [0,1], fp16. The midpoint (t = 0.5) is built in.
  • g16: fp16, except the full-size warp grid, which runs in int16 (an fp16 grid blurs by up to 0.3 px at these widths). On an S25 Ultra the 1280x736 build scores 63.5 dB against the ONNX model.

The weights are the authors', unchanged; only the graph was converted. MIT licence, as the original.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support