Geometry-calibrated.
Explicitly editable.
GeoCo-SAVi gives object slots readable position and scale, with predictable controls for moving, resizing, and exchanging appearance.
Explore the resultsInside an
object slot.
Each object is represented by appearance, position, and scale. GeoCo-SAVi calibrates these factors so their values align with decoded geometry and remain useful across appearances.
The visual content carried by an object slot.
An explicit location, aligned with the decoded center.
A readable extent, with a predictable resizing response.
A geometry-control formulation
Spatially equivariant decoding, calibrated through factual and counterfactual supervision.
Shared geometric meaning across appearances.
For video, a temporal initializer carries slot states into the next frame’s refinement.
Predictable edits,
across objects and video.
Position, scale, and appearance commands are evaluated alongside factual reconstruction and object grouping on Obj3D and MOVi-C.
Object editing on Obj3D
Single-frame editsEditing F1 (%) against ideal transformations of each model’s factual masks, evaluated on valid edits without boundary filtering.
Readable geometry.
Preserved reconstruction.
Obj3D · GeoCo-SAVi and ISALower position-to-decoded-center error
0.0107 GeoCo-SAVi / 0.1049 ISAReconstruction PSNR
44.210 GeoCo-SAVi / 44.145 ISA