A new preprint on arXiv (2610.15854) takes aim at a common assumption in feed-forward 3D reconstruction: that higher-quality input images lead to better scene geometry. The authors argue that for indoor scenes, grounded spatial reasoning relies on a geometric reconstruction that the multimodal system does not itself generate, and that this reconstruction is shaped more by which viewpoints are chosen than by how sharp each frame is.

The paper's title — "Sharper Frames, Worse Rooms" — makes the central claim explicit: viewpoint coverage dominates image quality. Because feed-forward multi-view reconstruction models work under a fixed view budget, the selection of frames becomes a decisive factor. The abstract, which is truncated in the listing, begins to explain this trade-off but does not yet detail the experimental setup or metrics.

As a preprint, the findings have not been peer-reviewed, and the full methodology is not available from the abstract alone. Still, the framing offers a useful caution for systems that naively prioritize image resolution over camera placement when reconstructing indoor scenes.