How Far Does a Shared Linear Map Go? Probing Feature-Space Manipulability for Image Editing
The research investigates if a single linear model operating on one feature vector can reconstruct images after complex semantic manipulations. It tests geometric, masking, and generative edits, finding that a simple linear map is sufficient for high-quality reconstruction. This suggests the feature space is organized in approximately linear structures where concept transformations correspond to subspace rotations and scalings.