Speridlabs Idol: Turn One Photo into a 3D Gaussian Splat, and How to Run It

Idol is an open image-to-3D model from Sperid Labs that turns a single photo into a 3D Gaussian splat, generating up to 1,048,576 Gaussians in under 2 seconds on a consumer GPU with 7.2 GB of VRAM. It runs locally through the idol-sdk Python package under Apache 2.0.

Sperid Labs released Idol on October 28, 2026, and it climbed into the top 20 of Hugging Face's overall trending list shortly after. The model reconstructs an object in 3D from a single picture, including sides the camera never captured, and writes the result as a standard .ply Gaussian splat file that opens in any 3DGS viewer.

Written by Priya Raghunathan, local-AI hardware reviewer. I verify parameter counts, licenses, hardware requirements, and every install command against the primary source before recommending a setup path.

A new top-20 entrant on Hugging Face's overall trending list

At time of writing the Idol model page shows 265 likes and 4,194 downloads, and the idol-sdk repository on GitHub has 441 stars. The most-liked Space built on it is the official Idol demo with 42 likes, a quick way to try the model on your own photo before installing anything.

Specs at a glance

Where the parameters and VRAM actually go

The 1.80B parameter figure covers six separate components, not one model. A diffusion transformer predicts the Gaussian positions, two decoders turn that prediction into an octree and then into full Gaussian attributes, and two borrowed encoders, BiRefNet and DINOv3, handle background removal and image understanding before any of that starts.

Add it up and weights alone run 3.7 to 4.4 GB, with the full request, including activations and the KV cache for sampling, peaking at 7.2 GB. That fits on most 8 GB consumer cards with room to spare.

Install and run locally

There is no separate weights download step. The first run pulls Idol's roughly 3 GB of weights and both image encoders straight into the Hugging Face cache.

Each run writes three files per image: a .ply Gaussian splat you can drop into any 3DGS viewer, a cropped and background-removed copy of the input, and a camera JSON file with the pose Idol predicted for the shot.

from idol import IdolPipeline

pipeline = IdolPipeline.from_pretrained("speridlabs/idol")
output = pipeline("photo.jpg", mode="real", seed=42, num_gaussians=1_048_576)
output.gaussian.save_ply("photo.ply")

Picking a mode and the useful flags

The --mode flag controls how tightly Idol crops the object out of the frame, and matters more than most of the other settings. Pick synthetic for renders and generated images of a single object, where the object already fills most of the frame. Pick real for photographs, which usually need more surrounding context to crop correctly.

How it compares to TripoSplat, TRELLIS and LGM

Sperid Labs benchmarked Idol against three other image-to-3D approaches, TripoSplat, TRELLIS and LGM, scoring the input view with PSNR, SSIM, LPIPS and IoU, and seven held-out new views with DINOv2, CLIP and a sharpness distance metric. All four models were measured on a single H100.

LGM still wins on raw PSNR and is by far the fastest of the four, but it falls well behind on the novel-view metrics, DINOv2 in particular, which suggests it reconstructs the visible side of an object well but guesses less convincingly at the sides the camera never saw. That is the gap Idol is built to close.

Sperid Labs also measured request time across ten different GPUs, same code and settings on each. It ranges from 3.4 seconds on a B300 down to 54.6 seconds on an L4, with an RTX 4090 landing at 19.1 seconds in between.

One image to 262,144 Gaussians in 3.4 s. Request time at 50 steps: median over 8 images, with the same code and settings on every GPU, one GPU per machine.

License and use

Idol ships under Apache 2.0, so commercial use, modification and redistribution are allowed without a separate agreement. The two gated encoders it depends on, DINOv3 and the FLUX.2 VAE, carry their own licenses on Hugging Face, so check those pages if you plan to ship a product built on top of Idol.

FAQ

What is speridlabs/idol?

Idol is an open-source image-to-3D model that converts a single photo into a 3D Gaussian splat, reconstructing the object including sides the camera never saw. It is built by Sperid Labs and released under Apache 2.0.

How much VRAM does Idol need?

7.2 GB of VRAM at peak, with 3.7 to 4.4 GB of that spent on model weights. It is designed to run on consumer GPUs rather than requiring a data center card.

How many parameters does Idol have?

About 1.80 billion parameters total, split across a 370.6M denoiser, a 50.8M octree decoder, a 237.3M Gaussian decoder, and two borrowed encoders, BiRefNet at 220.7M and DINOv3 ViT-H+/16 at 840.6M, plus an 84.1M FLUX.2 VAE.

Do I need a Hugging Face account to run it?

Yes. Idol depends on two gated Hugging Face models for image encoding, facebook/dinov3-vith16plus-pretrain-lvd1689m and the VAE from black-forest-labs/FLUX.2-dev. You need to accept both models' terms and log in with hf auth login before Idol will download anything.

Is Idol free for commercial use?

Yes, Idol itself is Apache 2.0, which permits commercial use, modification and redistribution without extra permission. The two gated encoders it relies on have their own separate license terms.

Where to go from here

• Try it before installing anything: the official Idol Space runs the full pipeline on an uploaded photo.
• Compare it with other local image and 3D models: the local AI model directory has specs side by side.
• Browse other recent how-to guides: the how-to index lists them newest first.