Constraint-Preserving Local Deformation of Radiance Fields via Anchor Ray Bundles

Abstract

The synthesis of novel views and the representation of complex three dimensional scenes have been revolutionized by implicit neural representations, most notably neural radiance fields. Despite their unprecedented ability to render high fidelity images from sparse inputs, the implicit nature of their underlying multi layer perceptrons poses significant challenges for intuitive scene editing and targeted local deformations. Modifying a specific region often leads to catastrophic forgetting or unintended global artifacts, as the spatial mapping is entangled within the network weights. Previous approaches have attempted to mitigate this by employing explicit proxies such as meshes or cages, but these methods frequently fail to preserve physical constraints, resulting in unrealistic stretching or distortion in adjacent unedited regions. In this paper, we propose a novel framework for constraint preserving local deformation of radiance fields using Anchor Ray Bundles. By defining specialized subsets of rays that act as spatial and photometric anchors between the deformed region and the static background, our method enforces geometric rigidity and visual continuity directly within the ray sampling process. We introduce a composite loss function that balances photometric reconstruction with a constraint penalty derived from these anchor bundles. Extensive experiments across multiple standard synthetic and real world datasets demonstrate that our approach achieves state of the art performance in maintaining structural integrity during local deformations. Our method significantly reduces artifact generation in unedited zones while allowing for complex, user defined geometric manipulations.