HelixBench: what spatial-acoustic consistency tests in AI video

HelixBench is the benchmark that comes with the HelixWorld paper. The arXiv abstract page says it formalizes spatial-acoustic consistency, meaning whether a generated sound field accurately tracks dynamic camera movement through 3D space. I read only the abstract page, so I do not quote scores or the test set.

What the benchmark is for

A video model with audio can fail this in plain ways: the sound stays centered while the camera pans, a source gets louder when the camera moves away, or the stereo image flips without a cause. The arXiv page says HelixWorld is conditioned on 6-DoF camera trajectories and user actions and synthesizes spatial stereo audio alongside the video, so camera motion is meant to show up in the sound.

A manual spot-check

You can spot-check any clip with headphones in a minute. Make a clip with a clear pan and one loud object, then listen for the questions below.

Where Sume fits

Sume's Gemini Omni Flash 1.1 renders native synced audio on every clip, and MiniMax H3 renders native stereo. Neither Sume catalog entry promises that the audio follows camera pose, so use the check above before you rely on it for a spatial-audio deliverable. The docs for both models cover resolution and duration, not acoustic geometry.