Skip to content

Varve optional AI models

Choose a tag to compare

@K-Arthur K-Arthur released this 14 Aug 02:26
· 7110 commits to master since this release

Optional AI model artifacts for Varve's Enhance workflow (downloaded on demand via the model manifest).

nafnet-gopro-width64-fp16b-embed.onnx — NAFNet-GoPro-width64 deblur

  • License: MIT (c) 2022 megvii-model (same as the official NAFNet repo)
  • Source checkpoint: NAFNet-GoPro-width64.pth, sha256 329d3ab4077b8d6b7ff61de376e483714667960bf85be027bf4335cda701196f (verified identical across independent mirrors)
  • Conversion: reproducible, see tools/nafnet-export/ in the repo (torch 2.13 CPU export, opset 18, dynamic HxW divisible by 16, fp16 internal weights with fp32 boundary)
  • Parity: bit-exact vs the trusted PyTorch reference on real photos (max diff <= 1/255, PSNR 94-96 dB); conversion fidelity 98.9 dB (fp32) / 65.2 dB (fp16) on the official GoPro test subset
  • GoPro test subset gate (24 images, seed 7): restored PSNR 32.20 dB mean (official 1111-image anchor: 33.71 dB)