Speridlabs Open-Sources Idol, a Model That Turns One Image Into a 3D Gaussian Splat

I have a soft spot for tools that turn one picture into something you can spin around, and Speridlabs just open-sourced a neat one. Their new model, Idol, takes a single image of an object and builds a 3D Gaussian splat from it, including the sides the picture never showed.

What it does

You feed Idol a photo, a 3D render, or an AI-generated image. It removes the background, frames the object, and outputs a standard 3DGS .ply file you can open in any Gaussian splat viewer. You can choose the level of detail, from 32,768 up to 1,048,576 Gaussians. On the project page, the team says a 262,144-Gaussian object is about 3.2 MB, while the full million-Gaussian version is around 12 MB.

The code is on GitHub and the weights are on Hugging Face under the Apache 2.0 license. The core model is about 0.66 billion parameters, and with its image encoders loaded, a full run peaks at about 7.2 GB of VRAM according to the model card, so this is meant for consumer GPUs. One catch: Idol reads images through two gated models, Meta's DINOv3 and the FLUX.2 VAE, so you need to accept their terms on Hugging Face first.

How fast and how good

Speridlabs' own benchmark compares Idol against TripoSplat, TRELLIS, and LGM. Idol leads most of the new-view quality scores, while LGM is faster but blurrier. Their timing table lists about 6.8 seconds per object on an H100 and about 19 seconds on an RTX 4090 at 262,144 Gaussians. These are the team's own numbers, so I'd treat them as a starting point.

Why I care as a fan: the gallery even includes an anime-style dragon, and I immediately thought of turning figure photos or character art into little 3D keepsakes for a shelf scene or a VR room. Just remember to use art you have the rights to. Details are on the Idol project page.