Every video call today works the same way: capture pixels, compress them, transmit them, and reconstruct them. But as video quality improves and immersive experiences become more common, moving millions of pixels across networks becomes increasingly costly.

Two recent research efforts explore that possibility through Generative Face Video Compression (GFVC). Instead of transmitting every frame, GFVC sends compact descriptions of facial expressions, motion, and semantics, allowing generative AI to reconstruct the video at the receiving end.

What if we stopped sending pixels altogether?

Generative compression systems often rely on different face representations, making interoperability difficult. The researchers introduced a framework that embeds compact facial information into existing video bitstreams using Supplemental Enhancement Information (SEI), allowing AI-driven video communication to work alongside established standards such as VVC, HEVC, and AVC. Notably, this work has already influenced emerging international standards for generative video coding.

Also, existing GFVC systems can distort backgrounds, causing flicker and visual instability. By separating foreground and background information before transmission and intelligently recombining them during reconstruction, the researchers produced more stable, natural-looking video while also enabling customizable backgrounds and chroma-key workflows.

We achieved 52.13% bitrate savings compared to VVC, perceptual quality improvements, and an 88.89% user preference score in subjective evaluations.

Despite challenges like reconstruction accuracy, decoder complexity, and latency, the future is promising. From ultra-low-bandwidth video calls to digital avatars and immersive virtual experiences, GFVC hints at a future where communication becomes smarter, lighter, and far more expressive.

Dolby's GFVC Papers:

Beyond Video Compression: Teaching AI to Reconstruct Human Presence

{{ pageEducationalSlot.headingContent[index] }}