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.