Title: DeblurNVS: Leveraging Geometric Latent Diffusion for Novel View Synthesis from Sparse Motion-Blurred Images

Abstract:

Novel view synthesis (NVS) remains a cornerstone challenge in computer vision and graphics. While recent innovations in generative view synthesis, 3D Gaussian Splatting (3DGS), and neural radiance fields (NeRF) have significantly elevated output quality, the majority of these techniques presuppose clean input data where cross-view geometric relationships and image structures remain intact. However, motion blur disrupts this ideal by obscuring local details and diminishing multi-view correspondences, a degradation frequently caused by camera instability, object movement, or limited shutter speeds during real-world capture. Although existing blur-aware NVS approaches attempt to mitigate this issue by simulating image formation processes, they are often hindered by the computational expense of per-scene optimization, which impedes scalable and generalizable sparse-view synthesis.

To overcome these limitations, we introduce DeblurNVS, a new framework designed to generate high-fidelity novel views directly from sparse, motion-blurred inputs without the need for scene-specific optimization. This method works by reconstructing the intermediate geometric representations essential for multi-view analysis, allowing blurred inputs to regain robust structural and correspondence cues. These recovered representations are subsequently integrated with target camera parameters to formulate the target-view representation, ultimately yielding a sharp RGB novel view. To facilitate large-scale model training, we developed a motion-blurred NVS dataset derived from DL3DV-10K, employing interpolation-based techniques to simulate finite-exposure blur. Comprehensive experiments confirm that DeblurNVS surpasses current baselines on synthetic motion-blur benchmarks and demonstrates strong generalization to real-world blurred scenes. The result is perceptually sharper and structurally more consistent novel views, achieved efficiently without expensive per-scene tuning.

Project page: https://github.com/PKU-YuanGroup/DeblurNVS

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC