Idol

Research checked 6 November 2026. Publisher claims and third-party trials are identified separately; this page is not an AIDreamHub hands-on benchmark.

One image to a splat, not to verified geometry

Idol, released by Speridlabs, takes an object photograph or render and generates a 3D Gaussian splat. Its standard output is a 3DGS .ply file, with up to 1,048,576 Gaussians and a default of 262,144. This is a representation for rendering and viewing, not automatically a polygon mesh, rigged game character, CAD model or manufacturing drawing.

The strongest near-term use is an exploratory product turntable, visual concept or interactive preview where generated unseen surfaces are acceptable. It cannot recover the true back of an object from evidence that does not exist. For an exact catalogue, authenticity or dimensional accuracy must come from additional photography, capture or modelling.

Official project page captured by AIDreamHub; not a locally run product.

Why preprocessing and the input camera matter

The pipeline removes the background, crops the object, encodes the image with DINOv3 and the FLUX.2 VAE, then generates and decodes the splat. The SDK distinguishes real photographs from synthetic renders because their crop margins differ. A poor cutout or wrong crop mode can become a bad 3D shape before sampling settings are even relevant.

Save all three outputs: the .ply, the preprocessed _input.png and _camera.json. Compare a render from the predicted input camera with the processed image before judging new views. If those do not align, inspect the mask and framing first. More Gaussians are not a substitute for a correctly prepared input.

How to use the official SDK

The documented baseline is Python 3.10+, PyTorch 2.5+ and an NVIDIA CUDA GPU. First accept access conditions for facebook/dinov3-vith16plus-pretrain-lvd1689m and black-forest-labs/FLUX.2-dev on Hugging Face, then authenticate locally. Model and encoder downloads occur on first use. Never put a Hugging Face token in a public prompt or commit.

After inspecting the repository, install into an isolated environment and run the photo example below with an image you own. Use --mode synthetic for renders instead. The directory example accepts multiple images, but that is not a production batch queue: retries, per-item status and output validation still need to be built. We did not execute GPU inference for this editorial review.

git clone https://github.com/speridlabs/idol-sdk.git
cd idol-sdk
python -m venv .venv
source .venv/bin/activate
pip install -e .
hf auth login
python -m idol.sample --image photo.jpg --mode real
python -m idol.sample --image renders/ --mode synthetic --out outputs

Community experience is more useful than a speed slogan

The official card’s headline says it can run in under two seconds, yet its separate cross-GPU table lists 19.1 seconds on RTX 4090 at 50 steps and 262,144 Gaussians. In a 3 November hands-on ComfyUI article, TNSOR_WORKS reported about 16 seconds without the turntable video and about 44 seconds with it on RTX 5090. These are different workflows and settings, not a contradiction that one universal number can resolve.

The same reviewer found a concave inferred back on a clock whose front looked convincing. Nynxz’s community integration also documents haze from low-opacity Gaussians and provides a prune node. Those observations identify things to inspect; they are not proof that every object fails or that pruning always preserves all details. We have not reproduced either test.

ComfyUI and delivery acceptance

ComfyUI-Idol is a third-party integration, not the official SDK. Its README uses stock TripoSplat nodes plus preprocessing, camera and prune helpers. Its sampling guidance maps SDK cfg_strength 3.0 to ComfyUI CFG 4.0, so copying a number between interfaces can change the experiment. Check the current node requirements rather than relying on a version mentioned in an older tutorial.

For a product pilot, review front, side, back and top views; reject changed logos, handles, openings and silhouettes. Then test the exported file in the actual web viewer or rendering application. A .ply extension alone does not prove every viewer supports the same Gaussian attributes. Mobile load time and memory deserve a separate delivery check.

Idol versus mesh generation or multi-view capture

Choose Idol when the deliverable is a visual splat preview and speed of iteration matters more than editable topology. Compare a mesh-oriented pipeline when you need UVs, rigging, collision geometry or downstream mesh editing. Compare photogrammetry or manual modelling when fidelity to a real object is the actual requirement.

The publisher compares Idol with TripoSplat, TRELLIS and LGM, but those results should not be read as a universal winner across formats or production tasks. In particular, a favourable image-similarity score cannot establish that a generated product has the correct hidden components. Define the required asset format before comparing quality.

Licenses and the realistic service opportunity

Idol and the community node pack are labelled Apache-2.0; the DINOv3 and FLUX.2 components have separate access conditions and licenses. The top-level model license does not erase those dependency obligations or grant rights to a photographed person, character or product image.

A useful service would add input checks, per-object review, predictable viewer delivery and a failure report. Merely connecting the model to ComfyUI is not an unexplored opportunity: that integration already exists. Price the accepted asset and human cleanup, not just the fastest generation headline. The hand-drawn image explains speculative unseen surfaces and is not an Idol render.

Questions before adoption

Is the output ready for 3D printing or a game engine?

Not automatically. Idol produces a Gaussian splat, normally stored as 3DGS PLY. A mesh conversion, if available in another workflow, is a separate operation and needs topology, scale, materials and application-specific checks. A convincing turntable does not establish watertight or dimensionally correct geometry.

Can Idol reconstruct the exact back of my product?

No single photograph establishes unseen details. Idol generates a plausible completion, and a community hands-on example showed an incorrect back surface on a clock. Inspect all delivery views and use extra photos, capture or manual modelling when exact product fidelity matters.

How much GPU memory and time should I budget?

The official card reports 7.2 GB peak VRAM for its measured full request, but runtime, precision, compilation and Gaussian count affect requirements. Its RTX 4090 timing table is 19.1 seconds, not the headline under-two-second scenario. Benchmark your full workflow, including downloads, warm-up, export and turntable rendering.

Sources reviewed

• Speridlabs — Idol model card and benchmark conditions
• Speridlabs — official SDK
• Nynxz — third-party ComfyUI integration
• TNSOR_WORKS — hands-on ComfyUI / Blender report, 2026-11-03

Explore related tools

• 3D tools
• Image tools