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

Abstract:

The advent of generative models has advanced multi-view image editing, inching us closer to the goal of broad 3D content generation and customization. However, the majority of current techniques rely on the geometry of the original, unedited scene to facilitate rigid modifications or changes limited to surface appearance. This dependency inherently restricts these methods to edits that maintain the underlying structural integrity of the scene. While other approaches have been specialized for particular tasks like adding or removing objects, general nonrigid edits—those that fundamentally alter scene geometry—remain a significant challenge for existing methodologies.

To address this gap, we introduce GeM-NR, a rapid, flexible, and training-free framework designed for general multi-view consistent image editing. This approach accommodates alterations that dramatically shift both the geometry and appearance of a scene. GeM-NR operates by taking a query unedited image and an anchor image (which has been edited using a selected backbone editor such as FLUX, Qwen, or BrushNet), ensuring the query image is modified in alignment with the anchor’s changes.

The methodology unfolds through three primary stages: 1. Depth Map Estimation: We employ a novel strategy aimed at maximizing the alignment between the 3D point clouds of the edited and unedited scenes. 2. Projection: The aligned data is projected onto the query viewpoint. 3. Refinement: The resulting image is refined based on the unedited query image.

This conditioning-based architecture demonstrates strong scalability, effectively managing transitions from two views to many views of a single object. Our results highlight GeM-NR’s capacity to manage edits involving substantial geometric and appearance shifts, a capability where prior methods often fall short. Comprehensive evaluations confirm that our approach enhances consistency across a broad spectrum of editing tasks, including the creation of 3D representations of the modified scene. Both quantitative metrics and qualitative assessments underscore GeM-NR’s state-of-the-art performance regarding edit fidelity, as well as its geometric and photometric consistency across multiple viewpoints.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC