GeoEdit: Geometry-Aware Object Editing via Dual-Branch Denoising

2026-06-29

AI Summary

This work addresses the challenge of single-image object geometry editing, which often suffers from perspective errors, ghosting artifacts, and background distortions due to insufficient 3D awareness. To overcome these limitations, the authors propose GeoEdit—a training-free Lift-Manipulate-Render-Denoise framework that disentangles foreground and background in 3D space. By aligning point correspondences and rendering geometrically consistent proxies using structural depth maps, GeoEdit preserves object identity while enabling free-form background synthesis. A key innovation is a variance-homogenized 3D constraint injection strategy that operates within a narrow noise-variance window, effectively preventing self-attention leakage. The study also introduces GeoEditBench, the first pose-aware benchmark for evaluating geometric editing fidelity. Experiments demonstrate that GeoEdit significantly outperforms existing methods in geometric accuracy, identity preservation, and background quality, particularly excelling in translation, rotation, and camera motion tasks.

Abstract

Precisely manipulating objects in a single photograph (translation, rotation, scaling) while obeying 3D physical constraints remains unsolved for diffusion-based editors. Current 2D methods lack spatial awareness and produce perspective violations. Forcing structural proxies into the latent space also disrupts variance homogeneity, and the resulting self-attention leakage leads to ghosting and background blur. The core difficulty is asymmetric: the relocated object must follow a rigid geometry, yet the uncovered background needs freedom to synthesize plausible content. We present GeoEdit, a training-free Lift-Manipulate-Render-Denoise pipeline that satisfies both constraints. We decouple scene and object in 3D, align them through point correspondence, and render a geometry-aligned proxy with a structural depth map. A Dual-Branch Denoising stage then refines this proxy: a video diffusion backbone preserves object identity, while 3D constraints are injected into the foreground within a narrow denoising window at matching noise variance (variance-homogeneous injection). The background denoises freely. Because the injected signal matches the native latent statistics, self-attention stays undisturbed. We also introduce GeoEditBench, a pose-aware benchmark covering object translation, object rotation, and camera movement with pose-aware evaluation metrics. Experiments confirm consistent gains in geometric accuracy, identity fidelity, and background quality. Our codes are available at https://github.com/Heey731/GeoEdit.

Problem

• object editing
• 3D geometry
• diffusion models
• spatial awareness
• perspective consistency

Innovation

• Geometry-Aware Editing
• Dual-Branch Denoising
• Variance-Homogeneous Injection
• 3D Object Manipulation
• Training-Free Diffusion Editing