Camera-conditioned video diffusion models generate video from text and camera parameters, but they frequently produce output that drifts from the intended camera path and shows geometric inconsistencies. According to the abstract of a new arXiv preprint, the root cause is that camera constraints and estimated geometry are only weakly enforced during the denoising process.
The paper, titled "GeoART: Geometry-Aware Refinement via Test-Time Scaling for Camera-Conditioned Video Diffusion," proposes a refinement strategy that applies geometry-aware adjustments at test time. The approach is framed as a form of test-time scaling, suggesting that additional computation during inference is used to better satisfy the camera and geometric constraints.
The source is a single arXiv announcement, and the abstract is cut off mid-sentence after mentioning "Existing remedies…". As a result, the specific techniques used by GeoART, how it compares with prior work, and any quantitative results are not available from the provided text. Readers should treat this as an early-stage description of the problem and proposed direction rather than a validated method.