Joint video-audio generation models have made notable progress in semantic correspondence and temporal synchronization, but the paper argues that applications such as AR/VR and interactive gaming demand something more: stereo audio that delivers an immersive sense of space. This requirement, the authors note, has been largely overlooked.
To address it, the researchers define a property they call Dynamic Spatial Correspondence — the perceived location of a sound should evolve consistently with the motion of its corresponding visual source. They propose StereoBind, a framework that binds visual source motion to stereo sound generation.
StereoBind coordinates visual motion and stereo audio through three complementary mechanisms: Visual Motion Binding, which establishes source-aware audiovisual correspondence; the Spatial Track Encoder, which captures absolute source positions; and Residual Track RoPE, which models relative motion.
For supervision and evaluation, the team constructed StereoWorld-29K, described as a large-scale stereo audio-video dataset with paired motion tracks, and StereoWorldBench, a benchmark for measuring audiovisual spatial consistency.
According to the paper, experiments show that StereoBind substantially improves spatial alignment in stereo audio generation compared with existing models while preserving overall audiovisual quality.