Abstract: Highlights•We propose an image-to-3D Gaussian optimization method for better view consistency.•We guide unseen views with multi-view priors using SDS loss for consistent shape.•We obtain consistent fine details with multi-view pixel-level supervision.•We estimate task uncertainty through backpropagation to balance loss components.•We achieve better view consistency than state-of-the-art methods.
External IDs:dblp:journals/prl/SahinAECL25
Loading