Idol: Specs, License & How to Run

Idol is Sperid Labs’ single-image → 3D Gaussian splat checkpoint. Official hosts we opened on 2026-11-02: the ungated card speridlabs/idol (created 2026-10-28T15:55:17Z; last modified 2026-10-29T18:17:36Z; 299 likes / 4,194 downloads), config.json, github.com/speridlabs/idol-sdk (504 stars / 38 forks; Apache-2.0 LICENSE), the project page speridlabs.com/research/idol, and the Gradio Space speridlabs/idol (46 likes; runtime zero-a10g when opened). API gated was false. Hugging Face pipeline_tag is image-to-3d.

This identification page hangs off Meshy alternatives. That mapping is editorial for a local image-to-3D checkpoint. We have no sourced quality comparison against Meshy, Luma, Tripo, TRELLIS, or TripoSplat.

It is not Iris-3B (same lab, pixel-space text-to-image). T2I facts stay on Iris-3B vs Qwen-Image-2.1-Turbo and the Midjourney-class hub. It is also not Qwen3.5-4B-3D-Coder (community Three.js code, not a splat).

The card’s “including the sides the picture never showed,” “under 2 seconds,” and “best in each column” bench sentences stay vendor.

What we verified (2026-11-02)

Field	Value	Source
Host	huggingface.co/speridlabs/idol (not gated)	HF API
Created / last modified	2026-10-28T15:55:17Z / 2026-10-29T18:17:36Z	same
Likes / downloads	299 / 4,194	same
License tag	apache-2.0	HF cardData
LICENSE on the weight repo	404	HF siblings
GitHub LICENSE / API	Apache License 2.0 / Apache-2.0	speridlabs/idol-sdk
Pipeline tag	image-to-3d	HF API
Library name	idol	HF API
Params (card full request)	~1.80B	README component table
usedStorage	3,572,854,245 bytes (~3.33 GiB)	HF API
Peak VRAM (card)	7.2 GB	README
Output	3DGS .ply (32,768–1,048,576 Gaussians)	README
GitHub stars / forks	504 / 38	GitHub API
Official Space likes	46	HF Spaces API

Card overview (copied): one picture of an object (render, generated image, or photo) → background removal + crop → up to 1,048,576 Gaussians as a standard 3DGS .ply. Modes: synthetic (default; object fills 1/1.2 of the frame) and real (photographs; 1/1.7). Default --num_gaussians is 262,144; --steps 50; --cfg_strength 3.0.

How to run (from the official card, not tested)

Copied names only. forkalt has not loaded a checkpoint and has not written a .ply.

The README says accept terms on two gated encoder cards, then hf auth login (or HF_TOKEN):

• facebook/dinov3-vith16plus-pretrain-lvd1689m
• black-forest-labs/FLUX.2-dev (VAE only)

git clone https://github.com/speridlabs/idol-sdk.git
cd idol-sdk
pip install -e .          # Python 3.10+, PyTorch 2.5+, NVIDIA GPU + CUDA
python -m idol.sample --image photo.jpg --mode real

Python entry copied from the card: IdolPipeline.from_pretrained("speridlabs/idol"). First run downloads Idol weights (~3 GB from this repo) plus the two encoders into the Hugging Face cache. Those encoder licenses are not the Idol Apache-2.0 tag.

Vendor benches (self-reported)

The card’s table compares Idol, TripoSplat, TRELLIS, and LGM at the input view (PSNR / SSIM / LPIPS / IoU) and seven new views (DINOv2 / CLIP / sharpness), 1 × H100 per model. Copied Idol row only (50 steps, 262,144 Gaussians): PSNR 12.25, SSIM 0.726, LPIPS 0.213, IoU 0.862, request time 6.8 s. Those are self-reported. They are not a Meshy, Luma, or Tripo ranking.

The across-GPU table (same 262,144-Gaussian setting) lists RTX 5090 14.1 s, RTX 4090 19.1 s, and RTX 3080 46.2 s. Vendor, not our measurement.

License

The weight card is tagged apache-2.0. There is no LICENSE file on speridlabs/idol (404 on 2026-11-02). LICENSE on speridlabs/idol-sdk is Apache License 2.0. We record both. The gated DINOv3 and FLUX.2 encoder terms are separate. Not legal advice.

What we did not invent

• Peak VRAM and “under 2 seconds” are vendor. We have not measured a consumer GPU.
• We do not claim splat quality versus Meshy, Luma, Tripo, TRELLIS, TripoSplat, or LGM.
• Idol is not a mesh/.glb studio on the README we opened; output is 3DGS .ply.
• Iris-3B is a different Sperid Labs card (T2I).

Related: Meshy alternatives — Iris-3B (same lab, T2I) — Iris-3B vs Qwen-Image-2.1-Turbo — Midjourney-class alternatives — Qwen3.5-4B-3D-Coder (Three.js code, not a splat) — SolidWorks alternatives (parametric CAD, different job) — Runway alternatives (video, different job). Catalog: home. Corrections: contact.