Speridlabs released Idol on October 28: an open Idol image to 3D model that turns a single picture of an object into a 3D Gaussian splat, including the back and sides the picture never showed. The code and weights are Apache 2.0, the download is about 3 GB, and it peaks at 7.2 GB of VRAM, so it fits on the kind of GPU a lot of creators already own.
Why care? If you make product shots, game props, AR scenes or concept turnarounds, “one image in, a rotatable 3D object out” removes a whole afternoon of modeling for rough assets. You can try it right now in a free browser demo with no install. There is one catch worth knowing before you build a workflow on it, and we cover it below.
Idol at a Glance
- What it is: a single-image to 3D Gaussian splat model from Speridlabs. Output is a standard 3DGS
.plywith 32,768 to 1,048,576 Gaussians. - What’s free: everything. There’s a free Hugging Face demo, plus open weights and an Apache-2.0 SDK.
- Hardware: an NVIDIA GPU with CUDA. Peak VRAM is 7.2 GB and the weights are about 3 GB. Speridlabs reports 19 seconds per object on an RTX 4090 and 46 seconds on an RTX 3080.
- Cost: $0. There’s no paid tier.
- The catch: it outputs splats, not meshes. It also needs two gated encoders (Meta’s DINOv3 and the FLUX.2-dev VAE), and their licenses apply on top of Apache 2.0.
What Idol Actually Does
You give Idol a render, an AI-generated image or a photo of one object. It removes the background with BiRefNet, crops and frames the object, then a 370.6M-parameter diffusion transformer generates the 3D structure. Two decoders turn that into an octree and finally into Gaussians. Counting the image encoders, the whole request is about 1.8B parameters.
For each image you get three files. name.ply is the splat. name_input.png shows exactly what the model saw after cropping, which is handy for debugging bad results. name_camera.json is the predicted camera for the input view. That camera file matters if you want to line the 3D object up with your original shot in a composite.

Gaussian Splats vs. Meshes: Know What You’re Getting
Idol does not make a polygon mesh. A Gaussian splat is a cloud of up to a million soft, colored ellipsoids. It looks photographic when you orbit around it, and fuzzy materials like velvet, fur and foliage often look better than they would on a low-poly mesh. But you can’t drop a splat straight into a 3D printer slicer, and most game pipelines still expect meshes with UVs.
Where splats do work today: browser viewers and editors like the free SuperSplat editor, web embeds, AR previews and splat plugins for the major engines. If you need a mesh, look at the alternatives table below or our earlier write-up on Hi3D’s open mesh refinement.
How Good Is It? Vendor Benchmarks, Read Carefully
Speridlabs compares Idol with TripoSplat, TRELLIS and LGM on one H100 each. They score each model on the input view (PSNR, SSIM, LPIPS, IoU) and on seven new views (DINOv2, CLIP, sharpness). Idol leads on six of the seven quality columns. Its novel-view DINOv2 score is 0.781, against 0.754 for TRELLIS and 0.743 for TripoSplat.
Two caveats. First, Speridlabs ran these numbers itself, and as of today we haven’t seen an independent leaderboard rank Idol. Second, most of the margins are small. SSIM is 0.726 vs 0.716 and IoU is 0.862 vs 0.860, which you won’t see by eye. The novel-view numbers matter most, because they measure the hidden sides that single-image 3D usually gets wrong, and Idol’s lead is clearest there. Treat it as “competitive with TRELLIS at a third of the VRAM,” not “best in the world.”
Speed needs a closer look too. The model card says Idol “can run in under 2 seconds,” but its own 50-step table shows 3.4 s on a B300, 6.7 s on an H100, 19.1 s on an RTX 4090 and 46.2 s on an RTX 3080 at 262,144 Gaussians. The sub-2-second figure presumably comes from fewer steps on datacenter hardware. On a consumer card, expect somewhere between 20 seconds and a minute per object.

Idol vs. Other Free Image-to-3D Models
| Model | Output | VRAM (vendor figure) | License | Free demo? |
|---|---|---|---|---|
| Idol | 3D Gaussian splat (.ply) | 7.2 GB peak | Apache 2.0 (+ DINOv3 and FLUX.2-dev terms) | Yes, HF Space |
| TRELLIS.2-4B | Textured mesh (GLB, PBR) | 24 GB minimum | MIT | Community Spaces |
| Hunyuan3D-2.1 | Mesh + PBR texture | ~10 GB shape, ~29 GB with texture | Tencent Hunyuan Community (excludes EU, UK, South Korea) | Yes |
| Hi3D v1.0 | Refined high-res mesh | Up to 96 GB at max resolution | See our Hi3D post | Yes |
Try Idol Free Today
No install: the browser demo. Open the Idol Space on Hugging Face. Click an example such as the backpack or cupcake, or use the upload icon under “Input image.” Pick Synthetic for renders and AI images or Real photo for camera shots, then hit Generate. In our test the backpack example took about 20–25 seconds end to end on the free ZeroGPU queue, and the result opened in an orbitable viewer with a download button in the top-right corner. Free ZeroGPU time comes with a daily quota, and a free Hugging Face login gets you more than browsing anonymously.
Run it locally (NVIDIA GPU with 8 GB or more of VRAM, Python 3.10+, PyTorch 2.5+):
- Accept the gated terms for DINOv3 ViT-H+/16 and FLUX.2-dev on Hugging Face (Idol only downloads FLUX.2’s VAE), then run
hf auth login. git clone https://github.com/speridlabs/idol-sdk.git && cd idol-sdk && pip install -e .python -m idol.sample --image photo.jpg --mode real(use--mode syntheticfor renders or a folder of images).
The first run downloads about 3 GB of Idol weights plus the two encoders into your Hugging Face cache. If you haven’t accepted the gated terms, Idol stops before downloading anything and tells you what’s missing. The full flag list is in the model card, and the method write-up is on the Speridlabs project page.

Honest Trade-Offs
- NVIDIA only. The SDK requires CUDA, and there’s no official Mac or AMD path yet. Mac users should stick to the free demo.
- Licensing is layered. Idol’s code and weights are Apache 2.0. But the FLUX.2-dev repo it pulls the VAE from is under Black Forest Labs’ non-commercial license, and DINOv3 has its own Meta license. Read both before you sell what you make, and don’t assume “Apache 2.0” settles the question.
- Splats, not meshes. Converting to a clean, printable or rigged mesh is a separate job. For that, TRELLIS.2 or Hunyuan3D are the more direct route if you have the VRAM.
- One object per image. It’s built for isolated objects, not full scenes or people in environments.
- Brand-new release. It came out two days ago, the GitHub repo has under 100 stars, and no ComfyUI node exists yet. Expect rough edges.
Workflow Tips
- Generate a clean source image first. One object, plain background, three-quarter angle. A fast local image model like FLUX 3 Image Klein is perfect for this. Our Flux 3 Klein vs Flux 3 video shows which one to pick.
- Match the mode to the input.
syntheticlets the object fill more of the frame, whilerealleaves more margin for photos. The wrong one crops your object badly. - Check
_input.pngbefore blaming the model. Most bad results come from background removal clipping a handle or strap. - Start at 262,144 Gaussians, then go up. Iterate at the default, then rerun your keeper with
--num_gaussians 1048576and a fixed--seed. - Want a person orbit instead of an object? Use a different tool. Our Krea 2 Anygles guide covers 360° views of people from a single image.
FAQ
Is Idol free?
Yes. The Hugging Face demo is free to use within ZeroGPU quotas, and the weights and SDK are free to download under Apache 2.0. There is no paid version.
Can I use Idol commercially?
Idol itself is Apache 2.0, which allows commercial use. It depends on gated DINOv3 and FLUX.2-dev components with their own licenses, though, so check those terms (and get legal advice if it matters) before using outputs in paid work.
What are Idol’s system requirements?
You need an NVIDIA GPU with CUDA and at least 8 GB of VRAM (7.2 GB peak), Python 3.10+, PyTorch 2.5+, and about 3 GB of disk for the weights plus the encoders. Speridlabs lists the RTX 3080 (10 GB) at 46 seconds per object.
Can Idol make a 3D mesh or STL for printing?
Not directly. It outputs a 3D Gaussian splat .ply. For meshes, use TRELLIS.2 or Hunyuan3D-2.1, or convert the splat with a separate tool.
For more hands-on comparisons of creative AI tools, subscribe to the Creative AI Show on YouTube and browse the rest of Creative AI Show.
Accurate as of October 30, 2026. Model versions, quotas and licenses change fast, so check the links above before you commit to a workflow.

No responses yet