Sperid Labs releases Idol, a 0.66B model that turns one image into 3D Gaussian splats in under 8 GB

Sperid Labs posted Idol on Hugging Face on 28 October. It is an image-to-3D model that turns one picture of an object into a 3D Gaussian splat, the point-cloud-like format that renders photoreal views in real time. The research page lists 0.66B parameters and output from 32,768 up to 1,048,576 Gaussians.

What it takes to run

The model card says peak VRAM is about 7.2 GB. A full request uses about 1.80B parameters once the image encoders are counted, with about 3 GB of weights. Sperid’s timing table, at 50 steps and 262,144 Gaussians, gives per-request times of:

GPU	Time per request
B300	3.4 s
H100	6.7 s
RTX 5090	14.1 s
RTX 4090	19.1 s
RTX 3080 10 GB	46.2 s

The card also says “under 2 seconds” elsewhere, which does not match its own table, so plan with the table. Code is in idol-sdk, and there is a Hugging Face Space demo.

Benchmarks and licence

Sperid compares Idol with TripoSplat, TRELLIS and LGM on its own benchmark. These are self-reported, and LGM beats Idol on PSNR in that table. The repository is tagged Apache-2.0 but has no LICENSE file, and that tag covers Idol’s own weights. Idol needs two other components: Meta’s gated facebook/dinov3-vith16plus encoder and the VAE from black-forest-labs/FLUX.2-dev, which come with their own licences. Check those before shipping a product.

Why this matters

An image-to-3D model with published weights that fits in under 8 GB, with timings down to a 10 GB RTX 3080, brings game-asset, product-shot and AR prototyping to ordinary desktop cards, while open interactive world models keep pushing 3D generation forward. The licence chain is the catch: an Apache tag on the main repo does not make the whole pipeline free to use commercially, since Black Forest Labs puts its own terms on FLUX weights, as with the Kommunity licence on FLUX 3 Image Klein.