Abstract
The demand for high-quality and immersive visual content continues to outpace the capacity of current 5G and future 6G networks, making compression an essential component of visual communication. Despite major advances in video coding over the past decades, key challenges such as latency, energy efficiency, scalability, and robustness remain unresolved. This talk will focus on one of these challenges: achieving effective compression at extremely low bitrates, where traditional codecs fail to preserve perceptual quality. Learning-based approaches enable substantial bandwidth reduction by exploiting the structure and semantics of visual content and can operate in a generative regime, where visual data are reconstructed by conditioning a trained model on compact latent or semantic representations. However, efficiently navigating the rate–distortion–perception trade-off with these models remains a major open problem.
I will illustrate these ideas through the examples of Generative Face Video Coding (GFVC), where realistic talking-face motion and texture can be synthesized from compact transmitted features, and generative 3D point cloud compression, where compact embeddings are used to guide a diffusion-based reconstruction. I will conclude by discussing how these concepts extend to semantic and task-oriented video communication, which generalizes traditional paradigms and opens new perspectives in this evolving field.
