Wang-pengfei / One2Scene

What is One2Scene?

One2Scene turns a single image into a geometrically consistent, explorable 3D scene you can freely navigate in 360 degrees, outputting videos, renders, and point clouds. Start with a Python script using HunyuanWorld to generate a panorama, feed it into the scaffold model for a 3D base via feed-forward Gaussian Splatting, then denoise for photoreal views at arbitrary cameras. This ICLR 2026 submission (check the github iclr 2026 openreview buzz) solves single-view 3D gen's distortion issues during big camera moves.

Why is it gaining traction?

Unlike prior methods crumbling on off-angle views, One2Scene delivers stable geometry via multi-view stereo priors and bidirectional fusion, topping benchmarks in panorama depth, 360 recon, and novel views per the arXiv paper. Devs dig the pretrained HF models (1.9GB scaffold, 19GB denoise) and demo pipeline yielding immersive outputs fast on GPUs. Amid iclr 2026 leak and reddit chatter on accepted papers, it's drawing eyes for practical 3D exploration without multi-view inputs.

Who should use this?

3D vision researchers benchmarking iclr 2026 papers or chasing SOTA scene gen. AR/VR prototype builders needing quick single-image worlds for apps. Gaussian Splatting fans tweaking configs for custom datasets like RE10K or ScanNet.

Verdict

Promising for iclr 2026 workshops with solid paper results, but 19 stars and 1.0% credibility signal early-stage research code—expect heavy setup (PyTorch 2.5, third-party deps like ZIM/Draco) and no tests. Fork and experiment if you're in 3D recon; skip for production.

(198 words)