GeoT2V-Bench: Benchmarking 3D Consistency in Text-to-Video Models via 3D Reconstruction

As text-to-video (T2V) models are increasingly used to generate virtual camera captures—such as orbiting objects or moving through rooms—it is no longer sufficient for videos to be merely visually plausible. This paper asks: when a model generates a video based on a camera-motion prompt, does that video provide coherent multi-view evidence for a single, rigid, static 3D scene?

The authors introduce GeoT2V-Bench, a diagnostic pipeline that treats generated videos as candidate 3D acquisitions. The pipeline follows a structured process:

1. Camera Estimation: It uses VGGT-style geometry estimation to derive per-frame camera intrinsics and poses from the generated video.
2. Flexible Fitting: It fits a Deformable Gaussian Splatting (DeformableGS) model to the video, which is intentionally over-parameterized to absorb temporal inconsistencies.
3. Static Proxy: It derives a 'MedianGS' static proxy by aggregating the temporal median of the flexible fit.
4. Diagnostic Profiling: Instead of a single score, it generates a continuous profile including apparent motion, static rendering error, flow agreement, and the gap between flexible and static fits.

By evaluating 3,840 reconstructions across 12 open-weight model configurations, the authors demonstrate that standard metrics for visual quality or temporal smoothness are insufficient for assessing 3D consistency. The study finds that visible motion, static rendering error, and flow agreement often disagree, meaning a video might appear smooth while failing to support a stable 3D reconstruction. The benchmark successfully identifies complementary failure modes, such as videos that rely on view-dependent repainting or nonphysical drift rather than representing a stable 3D environment.

This work shifts the evaluation of T2V models from 'does it look good?' to 'does it represent a 3D world?'. By providing a diagnostic profile rather than a pass/fail label, the benchmark helps researchers pinpoint exactly where a model fails—whether it is in camera trajectory estimation, static scene consistency, or the reliance on nonphysical deformations to hide geometric errors.