A new benchmark exposes how badly AI image generators fumble basic geometry across viewpoints, and offers a fix for catching it.
Researchers built DeformView, a wide-baseline multi-view image dataset with pixel-level annotations marking where geometric inconsistencies occur between generated viewpoints of the same scene. Using it, they tested existing multi-view consistency-scoring methods, tools originally built to evaluate novel view synthesis models, and found those methods transfer poorly to the forensic job of localizing exactly where geometry breaks down. In response, they built DEFECt3R, a lightweight classifier that reads cross-view feature relationships to flag inconsistent pixels, trained with hard negatives, geometrically consistent images that have still been deformed, to sharpen its judgment. Compared to prior consistency-scoring approaches, DEFECt3R cut false positives while improving localization accuracy.
The interesting part is the framing. Multi-view consistency has mostly been a quality metric for judging whether a 3D reconstruction model produced a convincing scene, not a forensic tool for catching manipulated images. Treating geometric mismatches as evidence, the same way file metadata or compression artifacts get treated today, gives investigators another signal for spotting AI-generated or altered photos when multiple viewpoints of a scene are available.
It is a narrow tool for a narrow slice of the deepfake problem, since it only works when you have more than one view of the same scene to compare, but a narrow tool that actually reduces false positives beats a broad one that cries wolf.