A method for reconstructing a complete, metrically scaled 3D scene mesh from a single photograph, built around an object-generator redesigned to handle whole scenes. The pipeline lifts the input image into a point map and splits space into adaptive 3D chunks that grow with distance to preserve nearby detail and cover distant structures. Image features are explicitly conditioned in 2D→3D by projecting them onto observed surfaces and marking voxels as free, visible, or hidden. Each chunk is produced by a sparse-structure transformer conditioned on those lifted features and visibility, followed by a structured-latent transformer; chunks are generated autoregressively with overlapping latents reused to enforce local consistency, and a decoder assembles chunks into a unified mesh. An interactive demo exposes many reconstructions and a distance-measure tool; displayed meshes are simplified for web viewing while full-resolution outputs are used for figures and videos.
The approach argues that adaptive chunking plus explicit visibility and autoregressive latent reuse yield high-fidelity, globally consistent reconstructions from one view. Results demonstrate plausible completion of occluded geometry, preservation of fine nearby detail alongside coherent distant structure, and visual improvements over contemporary single-image scene methods in comparisons. The system produces metrically meaningful meshes useful for visualization and downstream tasks where scene-scale geometry and consistency from a single photo matter.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Rgpu is a PyTorch device that allows tensors to reside on a remote GPU, enabling use from devices like Macs without CUDA. It simplifies remote GPU access for deep learning tasks. (github.com)
NanoMuse is an open-source AI agent designed to work across multiple devices, including phones, tablets, and computers. It remembers user preferences, stays active when apps are closed, and interacts through a personalized interface. (github.com)
OpenWAM is an open framework that combines video prediction and robot control using a shared causal video backbone. It supports various interaction models and improves robot task success rates through causal adaptation and counterfactual supervision. (openwam.stanford.edu)
OpenAI's Decisions API can evaluate GitHub pull requests by analyzing diffs against specific review questions, such as compatibility or migration needs. This helps organize review processes by providing clear answers to targeted concerns. (vercel.com)
Strands Decider 2B is an open-source decision model designed for building AI agents. It supports various tools, integrations, and deployment options for creating and managing intelligent systems. (strandsagents.com)
UniEvo-VL is a framework that enables multimodal models to improve themselves by using their own feedback during testing. It enhances image generation performance without external supervision, showing significant gains on benchmark tests. (arxiv.org)
Today's best Hacker News stories, summarized and screenshotted, one email a day.