Distilling codec-awareness into lightweight downscalers for real-time CPU deployment, with up to -8.77% BD-Rate PSNR gains across x264, x265 and VVenC.

Reinterpreting surrogate gradients as first-order local codec approximations, enabling full neural wrapping with up to -23.59% BD-Rate gains.

End-to-end training with real non-differentiable video codecs via surrogate gradients.