Grouping as Smoothing: Better Object Discovery Can Erase Local Appearance

Object-centric models are commonly ranked by segmentation quality, although grouping an object and preserving its content are different requirements. We test their relationship in Artificial Kuramoto Oscillatory Neurons (AKOrN), whose recurrent states live on a product of spheres. We replace the model's cross-token attention with a retrained per-token MLP matched to within 0.039% of its parameters. Across three training seeds, severance lowers canonical ID CLEVRTex FG-ARI from 0.757 to 0.386, but raises retrieval of held-out materials from 0.098 to 0.348 mAP; the paired gains are 0.247, 0.254, and 0.250. Ground-truth masks are applied only after each forward pass, so the retrieval gap cannot be attributed to different predicted masks. A same-activation OOD sweep gives the corresponding dose response: after two recurrent steps, FG-ARI and material retrieval are negatively ranked in all three checkpoints. Formally, the implemented recurrence is a retracted tangent-vector-field step on a product of spheres, with an exact decomposition into coupling and stimulus drives. The learned coupling drive is large and locally smooth, consistent with a symmetric-consensus model that locally attenuates higher graph-frequency modes. Thus, in this setting, better object discovery comes with substantially lower direct recoverability of local material identity from the evaluated representation.