T-GVC: Trajectory-Guided Generative Video Coding at Ultra-Low Bitrates

T-GVC, or Trajectory-Guided Generative Video Coding, represents a significant advancement in video coding technology, particularly for ultra-low bitrate scenarios. Traditional methods often struggle with domain specificity and high-level text dependencies, leading to unrealistic video reconstructions. T-GVC overcomes these challenges by integrating a semantic-aware sparse motion sampling pipeline, which captures pixel-wise motion based on semantic importance. This innovation allows for a substantial reduction in bitrate while maintaining essential temporal semantic information. Experimental results indicate that T-GVC outperforms both traditional and neural video codecs under ultra-low bitrate conditions, achieving more precise motion control than existing text-guided methods. The introduction of trajectory-aligned loss constraints further enhances the framework's ability to ensure physically plausible motion patterns, marking a notable leap in the field of generative video coding.