NVRC++ video codec runs 7.6x faster with high scalability

Implicit neural representations (INRs) have shown promise for video compression, but existing codecs struggle to scale across different bitrates and quality levels without sacrificing complexity. Lightweight models degrade at high quality, while high-performance models become too heavy. To solve this, the team introduces NVRC++, a novel INR-based video codec that uses a lightweight INR with multiple high-resolution feature grids. This design maintains consistent complexity across a wide range of bitrates, from 7k MACs/pixel to 360k MACs/pixel, all while supporting real-time decoding.

NVRC++ pairs this architecture with an optimization framework that efficiently overfits high-resolution grids for long video sequences, exploiting spatio-temporal redundancies without prohibitive computational or memory overhead. An advanced entropy model is specifically designed to compress the high-dimensional grid parameters, enabling significant bitrate savings. The experimental results show that NVRC++ delivers up to 7.6x faster decoding than the previous state-of-the-art INR codec (NVRC), while providing comparable or better rate-distortion performance.

This breakthrough could enable practical deployment of neural video compression on a wider range of devices, from low-power edge hardware to high-end servers. By unifying complexity and scalability within a single architecture, NVRC++ addresses a key limitation that has prevented INR-based codecs from entering real-world applications. The paper is available on arXiv (2606.28163) and represents a significant step toward efficient, scalable neural video compression.