Uncertainty-Adaptive Noise Scheduling for Robust Refinement
Map geometric warping uncertainty scores to continuous noise injection levels so diffusion models can properly regenerate severely distorted image regions.
Source paper
Generated by AI. The paper’s authors did not write these.
Map geometric warping uncertainty scores to continuous noise injection levels so diffusion models can properly regenerate severely distorted image regions.
Replace transient local 3D Gaussian caches with a persistent global map to retrieve valid geometry when cameras return to previously visited locations.
Integrate a motion segmentation module into the cache creation pipeline to treat moving objects as occluded and force context-based diffusion regeneration.
This paper presents WorldWarp, a framework for long-range, geometrically consistent novel view extrapolation from limited input images. The approach combines an online-optimized 3D Gaussian Splatting (3DGS) cache, which provides structural geometric priors via forward warping, with a spatially-temporal diffusion model (ST-Diff) that refines these priors using a "fill-and-revise" noise schedule to handle occlusions and correct artifacts.
The method operates autoregressively, generating video chunk-by-chunk. For each chunk, it optimizes a local 3DGS representation on recent history frames to render warped images for the next camera trajectory. These warped images act as dense conditionings for a non-causal video diffusion model, which is trained with a spatially-varying noise schedule: valid warped pixels receive low noise (for refinement) and invalid/occluded pixels receive high noise (for generative hallucination).