LibreYOLO Adds Python Support for Sperildabs Idol, Which Creates 3DGS From a Single Image

What happened

The documentation for LibreYOLO, a collection of open-source vision models, now has a page for Speridlabs Idol. Given a single image of an object, the model outputs the whole object as 3D Gaussian Splats (3DGS). Output is in the standard .ply format or the web-oriented .splat format.

How it works

To use it, install libreyolo[gen] with pip, pass a model name to the LibreGen class, and feed it an image. Accepted inputs include paths, URLs, PIL images, NumPy arrays and folders. An image with transparency is used as-is as a cutout; otherwise, background removal (BiRefNet) runs first. There is no text input, and the resolution is fixed.

By the numbers

Why it matters

3DGS is normally reconstructed from many photos. If it can be generated from one, it could widen the entry point for making assets for games and exhibitions. However, this is generation that estimates surfaces that are not visible. It differs in nature from recording uses that measure a real object, so accuracy cannot be assessed from the available material. Because output is available as .plyand .splat, the files can be loaded into existing 3DGS viewers, which should make it easy to handle in practice.

Examples

• The page includes examples generated from 11 photos of sculptures and architecture in the Basque Country, such as the Guggenheim Museum Bilbao and Maman. The examples are shared under CC BY-SA 4.0 and compressed to about 4MB each.

How to try it

1. Install the dependencies with pip install "libreyolo[gen]"
2. Load the model with LibreGen("speridlabs-idol") (the first run downloads about 5GB)
3. Pass an image and save with result.save("out.ply") or a .splat file
4. Use mode="real" for photos, steps=25 to prioritize speed, and increase gaussians to prioritize detail

Caveats and open questions

• Quantitative evaluation of generation quality or shape accuracy cannot be confirmed from the available material
• Export to ONNX, TensorRT and similar formats is not supported. train() and val() are also not implemented and raise NotImplementedError
• The M4 Pro measurement covers only two objects and is very slow, at about 460 seconds
• Even with the same seed, CUDA, MPS and CPU produce different objects. Even on CUDA, the documentation says the positions of some Gaussians occasionally change from run to run
• The DINOv3 encoder is under Meta's own license, which has use restrictions such as military and weapons applications. The license of each set of weights must be checked before commercial use
• The source is the library's documentation. An evaluation from Speridlabs itself cannot be confirmed within the available material
• The X post that prompted this topic was a summary by a search AI. The facts in this article are written from the documentation

Glossary

3D Gaussian Splatting (3DGS)

A technique that represents a scene with many semi-transparent 3D Gaussian distributions and renders it quickly.

.splat

A lightweight 3DGS format for web display. This page gives 32 bytes per Gaussian.

classifier-free guidance

A technique for adjusting how closely a generative model follows its conditions.

Sources

Primary sources (publisher, paper, repository)

• Speridlabs Idol in Python: one image to 3D Gaussian splatsLibreYOLO（libreyolo.com）・en

The X post that surfaced this story

• View @EHxuban11's post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original