TL;DR of the paper, generated automatically.

NeuOctree is a neural shape representation that can be ray traced directly, with no mesh extraction or texture baking step.

• How it works: Features are stored at the corners of sparse octree voxels and interpolated into a continuous field. A lightweight decoder with recurrent intersection-guided sampling refines each ray query near the predicted surface.
• Results: Across seven meshes and four storage budgets, the paper reports the lowest positional error in 24 of 28 comparisons, with competitive normal accuracy.
• Speed and transfer: The feedforward model builds a representation in 0.05 seconds from colored point clouds, and it transfers from Objaverse to ShapeNet without retraining.
• Scope: It handles open, non-watertight geometry, and the paper also demonstrates ray-traced rendering and image-conditioned generation. Code is linked from the abstract.