Research & Papers

DCVC-MB Uses State Space Models to Cut Video Bitrate by 30%

New neural codec beats H.266/VVC by 30% with Mamba architecture and entropy skipping

Deep Dive

DCVC-MB (DCVC-Mamba) introduces a spatio-temporal fusion model based on state-space models (SSMs) for bidirectional temporal prediction in B-frame coding. The framework uses an IBP (Intra-Bidirectional-Predictive) frame strategy for low-delay scenarios, combining past and future reference frames to improve compression efficiency. A key innovation is an entropy-aware skipping mechanism that selectively omits coding certain latents, significantly reducing entropy coding time without sacrificing quality. The model also incorporates two inference-time strategies to further boost performance.

In experimental evaluations, DCVC-MB achieves average BD-rate reductions of 8.98% compared to existing neural video codecs. Against traditional standards, it shows 30.45% improvement over VTM-19.0 in low-delay P configuration and 1.81% over VTM-19.0 in random-access mode with Inter-GoP=16. The work, accepted at ICME 2026, demonstrates that state-space models can effectively replace transformers and RNNs for video compression, offering a new path for efficient neural codecs.

Key Points
  • DCVC-MB uses state-space models (Mamba) for spatio-temporal fusion, achieving 8.98% BD-rate reduction over prior neural video codecs.
  • Outperforms traditional H.266/VVC benchmark VTM-19.0 by up to 30.45% in low-delay mode and 1.81% in random-access mode.
  • Entropy-aware skipping mechanism reduces coding time by selectively omitting certain latents during entropy encoding.

Why It Matters

Cuts video bandwidth by 30% without quality loss, enabling cheaper streaming and storage for professionals.

📬 Get the top 10 AI stories daily