Idol

speridlabs.com · listed 02 Nov 2026

Turn an object photo into a Gaussian splat with a local Python workflow

Design Image generationOpen sourceCLIOpen source

Visit Idol See alternatives (6)

PREVIEWfrom speridlabs.com · 02 Nov 2026

Verified 02 Nov 2026

✓ Online · checked 02 Nov 2026✓ Reviewed by an editorHow we check

Overview

What is Idol?

Idol is a Speridlabs research model for reconstructing an object from one image. Its Python package isolates the foreground, prepares the crop and produces a representation that Gaussian-splat viewers can rotate. The result is stored in a PLY file alongside the prepared input and an estimated camera configuration. Users can choose the splat count and sampling settings through Python or a command line. The SDK is licensed under Apache 2.0, and the published Idol weight card declares the same licence. Running it still needs computing resources and access to separately governed image encoders. DINOv3 retains its own agreement; Black Forest Labs licenses its FLUX.2 autoencoder under Apache 2.0, separately from the full generator. This entry covers the downloadable project at speridlabs.com/research/idol, not every Speridlabs product. Documentation and publisher artwork were inspected; no generation or performance test was run.

Good for

Developers prototyping image-to-3D workflows that consume Gaussian splats
Researchers comparing single-view object reconstruction with configurable sampling

Watch out for

A splat PLY represents visual appearance; the documented output does not promise a polygon mesh, animation rig or CAD geometry. Surfaces hidden in the input are inferred and may be inaccurate.
The zero-dollar plan covers downloadable software and declared Idol weight licensing. GPU time, electricity and any rented hardware remain your expense; no managed API credits or ongoing free hosted allowance are included.
Setup requires a Hugging Face account with permission to retrieve both image encoders. DINOv3 has separate usage and redistribution obligations; do not describe the entire dependency chain as Apache licensed or universally unrestricted.
The publisher reports roughly 7.2 GB peak GPU memory, but speed varies with hardware and settings. The README headline and benchmark tables give different timings; aidirecto has reproduced neither.
The displayed image is an official social preview. It is publisher artwork, not a screenshot captured during an independent product trial.

Features

What it does

01Image preparationThe workflow removes the background and changes its crop for photographs versus synthetic source images.
02Adjustable splat densitySelect a Gaussian count between 32,768 and 1,048,576 to trade detail against representation size.
03Reusable output filesEach run can save the splat, the processed input image and estimated camera data for inspection in downstream applications.
04Python and terminal entry pointsCall the pipeline from application code or invoke its command-line module for a single image or a folder.
05Sampling controlsSet the seed, step count, guidance and optional compilation behaviour when exploring reconstruction quality and runtime.

Pricing

Idol pricing

From the product site · 02 Nov 2026

Self-hosted SDK and Idol weights$0No licence charge for the Apache-licensed SDK and the weights marked Apache 2.0 on the Idol model card. Computing resources are excluded. Separate encoder terms and gated download approval apply; this is not a hosted service subscription.

Crawler record

Change log

What our crawler noticed on speridlabs.com since listing.

02 Nov 2026+Listed after editor review

FAQ

Questions about Idol

What does the free plan cover?

The downloadable SDK has an Apache 2.0 licence and the Idol model card declares Apache 2.0 for its weights. There is no software subscription in this recorded offer. You supply the hardware, and external components keep their own terms.

Does it download the complete FLUX.2 image generator?

The inspected encoder code loads its VAE component. Black Forest Labs explicitly assigns Apache 2.0 to that autoencoder, separately from the full FLUX.2 dev model. Download access still requires approval on the gated repository.

What is required to run the documented workflow?

The setup guide specifies Python 3.10 or later, PyTorch 2.5 or later and an NVIDIA CUDA GPU. It also asks users to obtain access to the DINOv3 and FLUX.2 repositories before the first run.

Is the generated object an accurate reconstruction of unseen geometry?

Only one view is supplied, so unseen surfaces are generated hypotheses. The project exports Gaussian splats; assess visual fidelity and compatibility for your intended use rather than treating the output as measured geometry.