Codec-aware learning framework enabling end-to-end training of neural downscalers through non-differentiable video codecs via surrogate gradients. 6.11% BD-Rate improvement over codec-agnostic approaches.