SmoothMotionVectors: Optimizing Your Content for Video Codecs in Free View Video Compression

Disney Research Studios is pushing free-view video compression closer to something production teams can actually use. In work presented for SIGGRAPH 2026, the group studied how common pipeline choices affect 3D Gaussian Splatting reconstructions of dynamic scenes, with a focus on storage efficiency and fidelity rather than inventing a new scene representation.

The practical takeaway is that temporal chunking matters. Splitting long sequences into shorter segments, such as GOPs, makes optimization easier and can improve reconstruction quality, though it also adds some storage overhead. That tradeoff will sound familiar to anyone balancing encode complexity, asset size, and runtime quality in content-heavy pipelines.

The bigger result comes from shaping deformation signals so they produce smooth motion vectors across space and time. That makes the data friendlier to standard video codecs, improving rate-distortion performance and helping the codec do more of the heavy lifting. Across multiple datasets, the unified pipeline reportedly cuts storage by 20% versus state-of-the-art methods while maintaining or improving visual quality.

For developers working with volumetric capture, free-view video, or any codec-driven reconstruction workflow, the message is straightforward: preprocessing and training choices can matter as much as the representation itself. The team also calls out variability and ambiguity in current training and evaluation practices, which suggests there is still room for the field to converge on more reliable benchmarks.