InnVC neural video codec slashes bitrate by 46% at high fidelity

Learning-based video compression has made strides, but most methods rely on non-invertible transforms that introduce distortion even when quantization errors are small. At high-quality regimes, this transform error dominates, limiting fidelity. To solve this, Teng et al. introduce InnVC — an invertible neural network video codec. The core innovation is preserving an invertible main transform path before quantization, while feeding content-adaptive context via a compact implicit conditioning field. This separates strongly correlated video content from harder-to-model fine details, letting each component specialize. Additionally, a scheduled masking strategy progressively concentrates informative content into fewer latent channels, improving entropy coding efficiency.

On the UVG and MCL-JCV datasets, InnVC delivers strong performance across a broad quality range, excelling in high-fidelity compression. It yields BD-rate reductions of 21.66% in PSNR and 46.06% in MS-SSIM relative to the x265 codec on UVG. Notably, InnVC is the first neural video codec to cover operating points from low bitrate (around 20 dB PSNR) to high fidelity (over 40 dB) within a single architecture scale. This means one model can handle everything from bandwidth-constrained streaming to archival-quality compression, potentially replacing multiple codec profiles. The work is available as arXiv:2606.13957.