Idol is an image-to-3D model from Speridlabs, released at the end of October 2026 and at the top of Hugging Face's trending list since. Give it one picture of an object and it returns a 3D Gaussian splat of up to a million Gaussians, including the sides the picture never showed. It is released under Apache 2.0, and ComfyUI runs it natively, with its own template.
We ran that template on the smallest kind of GPU we rent and measured what it takes.
The input: one image, generated with Z-Image Turbo through our image API, 1024×1024, seed 4102. Prompt below.
What we measured
On an RTX 5060 Ti (16 GB) rented on ParalonCloud at $0.19 an hour, running ComfyUI's Idol: Image to Gaussian Splat template with its default settings:
| RTX 5060 Ti (16 GB) | |
|---|---|
| First run (loads the models) | 167 s |
| Every model after, splat + mesh + turntable video | 151 s |
| Peak VRAM | 5.6 GB |
| Cost per model at $0.19/h | about $0.008 |
| Models to download | about 3.3 GB (once per rental) |
| Output | Gaussian splat, a GLB mesh of 3.2 million triangles (75 MB), a 3 s turntable video |
Two things about these numbers:
- 151 seconds is the whole template, not the model alone. Besides the splat, the template turns it into a mesh and renders a 75-frame video, and those steps are a good part of the time. Speridlabs quotes "under 2 seconds" for the splat on its own on fast data-center GPUs; we did not measure the steps apart.
- 5.6 GB of VRAM, below the 7.2 GB Speridlabs gives for its own code. ComfyUI loads the fp16 files. Our ComfyUI rentals start at 16 GB, so on any of them Idol has room to spare.
What it costs, here and elsewhere
| Per model | What you get | |
|---|---|---|
| fal.ai hosted API | $0.10 | splat + point-cloud preview |
| ComfyUI on a rented RTX 5060 Ti | about $0.008 | splat + mesh + video, your own settings |
At $0.19 an hour, billed by the minute, a rented 5060 Ti makes about 24 models an hour. Even a single model, with the start of the rental and the model download, costs a few cents, so the rental is cheaper from the first model on. You also keep the workflow: the number of Gaussians, steps, guidance, seeds, and anything else ComfyUI can plug in before or after.
Run it in ComfyUI on a rented GPU
- Pick a GPU with 16 GB of VRAM or more (the RTX 5060 Ti is the cheapest), choose ComfyUI in the first step of Rent, and start it. Rent a GPU for ComfyUI walks through it.
- Open ComfyUI, then Templates, then Idol: Image to Gaussian Splat (in the 3D group).
- ComfyUI sees the four model files are missing and offers to download them, about 3.3 GB. They come from Comfy-Org/Idol, the same weights as Speridlabs' own in fp16, with the image encoders included, so no Hugging Face account or gated access is needed.
- Drop your image on Load Image and press Run. The first run takes longer while the models load. Change the seed for a new version of the same object; ComfyUI does not run again what has not changed.
Your splat, the GLB and the video land in ComfyUI's output folder, and you download them from the ComfyUI page.
Which pictures work
- One object, whole, with room around it. Idol reconstructs an object, not a scene.
- Generated images work best. A figure on a plain background, like the one above, comes out cleanest. You can make one with our image API for $0.003, or in ComfyUI on the same rental with Z-Image Turbo.
- Photos work too, with the background removed and the object framed; Speridlabs' own code has a separate setting for photos, which leaves more room around the object.
- Thin parts are hard: spokes, wires, fur standing out, glass.
The prompt of the input above:
Collectible vinyl toy figure of a cute capybara wearing a conical straw hat and a green
tech robe, full body, standing, three-quarter view, Pixar-style 3D character, single object
centered, whole figure visible with margin around it, plain pure white background, soft
even studio lighting, no shadow, no text, no logos
curl https://paraloncloud.com/v1/images/generations \
-H "Authorization: Bearer $PARALON_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "z-image-turbo", "prompt": "Collectible vinyl toy figure of a cute capybara ...", "size": "1024x1024", "seed": 4102}'
From the splat to a 3D print
The template's GLB is a real mesh, with colour on every vertex, and it has no holes. It is still not ready for a slicer. We checked the one from the capybara:
| Our GLB | |
|---|---|
| Triangles | 3.2 million (75 MB) |
| Open edges (holes) | 0 |
| Separate pieces | 80 |
| Edges shared by more than two faces | 58,021 |
| Units | none (about 1 × 1 × 0.7) |
The 80 pieces are shells lying on top of each other, the robe over the body and so on, and the edges shared by three or more faces are where they cross. A slicer can misread both. Five minutes in Blender fix it:
- Import the GLB.
- Add a Remesh modifier in Voxel mode with a voxel size of about 0.005 (half a millimetre on a 10 cm print) and apply it. It merges every piece into one closed shell.
- Add Decimate down to about 300,000 to 500,000 triangles, plenty for a printer.
- Scale it to the size you want in millimetres and export STL or 3MF.
Bambu Studio and PrusaSlicer repair what is left. On a one-colour printer only the shape matters; the vertex colours can go into a 3MF for a multi-colour printer, with less detail than on screen.
Good to know
- Licence. Idol is Apache 2.0, the weights and the code: what you make with it can be used commercially.
- The splat is the native output. The
.plyopens in any Gaussian splat viewer and in tools like SuperSplat; the mesh is a conversion of it. - Everything is deleted when the rental ends, the models too. Download your files before you stop it; next time, the 3.3 GB come back in a minute or two.
Ready? Pick a GPU and choose ComfyUI.
