Enhanced Multi-Resolution Generative Face Video Compression

This paper presents an advancement in Generative Face Video Compression (GFVC) recognized through contributions in recent JVET meetings. We optimized the general framework of GFVC by eliminating the need for an analysis model at the decoder side, and encoding the feature into bitstream at the encoder instead, which reduces the parameters and complexity of decoder. Furthermore, the paper discusses strides made in improving the architecture design of GFVC, these changes substantially reduce parameters and computational complexity. The proposed enhancements also focus on extending support for multi-resolution, as the GFVC framework was previously restrained to 256x256 video sequences. By ensuring scalability and adapting to evolving standards, our improved method demonstrates the potential for exceeding VVC standard performance, aligning with the trajectory of current technological standards and user expectations.