Title: Ultra-Fast Neural Video Compression

Abstract:

Although neural video codecs (NVCs) have proven their ability to deliver exceptional compression ratios, their extreme computational demands continue to hinder practical, real-world application. To address this challenge, we present a chunk-based coding framework that optimizes the balance between rate, distortion, and complexity. Rather than handling frames in a sequential manner, our method groups multiple frames into chunks, encoding them into a unified, compact latent representation for simultaneous decoding. This process relies on cross-frame interaction modules to perform joint spatial-temporal modeling, alongside frame-specific decoders that enable parallel reconstruction.

This new approach not only accelerates coding throughput significantly but also allows for more robust modeling of long-range temporal dependencies. To further maximize efficiency, we introduce a simplified entropy coding mechanism that merges bit-stream interactions into a single operation, thereby cutting down decoding overhead. Leveraging these advancements, we developed DCVC-UF (Ultra-Fast), an NVC that establishes a new state-of-the-art in performance. Experimental results demonstrate that DCVC-UF achieves exceptional encoding and decoding speeds, surpassing existing top-tier codecs. This work marks a significant milestone in the development of neural video compression. The source code is available at https://github.com/microsoft/DCVC.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC