GeM-NR: Geometry-Aware Multi-View Editing for Nonrigid Scene Changes

Researchers have introduced GeM-NR, a novel approach for multi-view consistent image editing. Unlike many existing methods that are limited to rigid or appearance-only modifications, GeM-NR is designed to handle nonrigid edits that involve substantial changes to a scene's geometry and appearance. The method operates in a training-free manner and can be integrated with various backbone editors such as FLUX, Qwen, and BrushNet. GeM-NR processes an edited anchor image and an unedited query image through several stages: depth map estimation with optimized 3D point cloud alignment, projection onto the query viewpoint, and refinement conditioned on the unedited query. This conditioning-based formulation is scalable for editing objects from two to many views. Evaluations indicate that GeM-NR improves consistency across a variety of edit tasks, including the generation of 3D representations of edited scenes, demonstrating state-of-the-art performance in edit quality and geometric and photometric consistency.