Idol-ComfyUI

Idol-ComfyUI

Turns one picture of an object into a full 3D Gaussian splat inside ComfyUI, for artists and developers who want 3D from a single image.

Made by PedroMarinhoDev on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Idol-ComfyUI holds the model files that let ComfyUI run Idol from speridlabs. Give it a single picture of an object and it produces a full 3D Gaussian splat of that object, capped at 1,048,576 Gaussians. The back and sides get filled in even when the input does not show them. It runs on ComfyUI's own nodes rather than a wrapper.

Who it suits: It suits 3D artists, game developers and ComfyUI users who want a 3D asset from one image and are comfortable installing custom nodes.

Industry | Who uses it | How they use it | Example
Games | 3D artists | Turn concept art of a prop into a 3D splat | An artist feeds in a picture of a treasure chest and turns it on a turntable
Product | Design teams | Preview an object in 3D from one photo | A team feeds in a photo of a vinyl figure and checks its back side
Research | ComfyUI and 3D researchers | Run Idol with ComfyUI's memory management | A researcher compares its output with the original SDK

If you need an editable mesh or have no GPU with about 6 GB of VRAM, you can skip it.

Where to use it:

• ComfyUI 0.43 or newer, on Windows, macOS or Linux
• ComfyUI-Manager, for installing the ComfyUI-Idol nodes
• A terminal with the hf command, for downloading the model files

Use case | Who | What they get back
Image to 3D splat | Artist or developer | A 3D Gaussian splat of the whole object
Turntable preview | Artist | A view that starts at the viewpoint of the input picture
Background removal on the input | Any user | The object cut out before the 3D step

How to set it up:

• ComfyUI 0.43 or newer
• A GPU with about 6 GB of VRAM, as the page's badge says
• Disk space for the model files, from 0.58 GB to 1.48 GB each depending on precision
• Python, for the hf command

1. Install the ComfyUI-Idol nodes. Use ComfyUI-Manager, run the comfy command below, or place a clone of the node pack in your custom_nodes folder.
2. Restart ComfyUI.
3. Download the model files with the hf commands below. The first hf line fetches Idol and the second fetches the helper files.
4. Open the template and run it.

pip install -U huggingface_hub
comfy node install comfyui-idol

cd ComfyUI/models
hf download PedroMarinhoDev/Idol-ComfyUI diffusion_models/idol_fp32.safetensors vae/idol_vae_decoder_fp32.safetensors --local-dir .
hf download VAST-AI/TripoSplat clip_vision/dino_v3_vit_h.safetensors vae/flux2-vae.safetensors background_removal/birefnet.safetensors --local-dir .

How to use it: In ComfyUI, open Templates, then ComfyUI-Idol, then idol_image_to_splat. Pick your picture in Load Image and run. The template already sets the sampler to euler with the simple scheduler, 50 steps and cfg 4.

Tool | What it does
Load Image | Picks the picture of your object
Load Diffusion Model | Loads the Idol denoiser file
KSampler | Runs the denoising steps with the preset values
TripoSplat Decode | Turns the result into the 3D Gaussian splat
Idol Input Camera | Starts the turntable at the input picture's viewpoint

Example: You pick a picture of a toadstool cottage in Load Image and run the template. The workflow returns a 3D Gaussian splat of the whole cottage, including the sides the picture does not show, which you can turn around.

For newcomers and experts: A Gaussian splat is a 3D shape built from many soft colored blobs. The easy first try is to install the nodes, download the files, open the template and run it with one clear picture of a single object.

For experts, the files hold the original Idol tensors. The fp32 files match them tensor for tensor, and the fp16 files are casts of them. The denoiser's stored token positions are left out because ComfyUI computes the same Sobol sequence, and the converter checks they are equal. The two decoders share one file. The page reports a relative error of 3.5e-6 for the denoiser and 9e-5 for the decoders against the reference SDK in fp32.

Your picture goes into Load Image
BiRefNet removes the background
DINOv3 and the FLUX.2 VAE read the image
The Idol denoiser runs in KSampler
TripoSplat Decode builds the Gaussian splat

Good to know:

• Extra files: the helper models DINOv3, the FLUX.2 VAE and BiRefNet come from the TripoSplat repository, the same ones its ComfyUI template uses
• fp16: the page says the fp16 denoiser gives the same speed and the same result at the default precision
• Output: a 3D Gaussian splat, at most 1,048,576 Gaussians
• License: Apache 2.0, the same as the original Idol weights
• Cost to run: none named, apart from a GPU with about 6 GB of VRAM
• Data it reaches: downloading uses the web, and the pictures you load stay in your ComfyUI
• Not for: ComfyUI older than 0.43, since the nodes need that version or newer

On Hugging Face, 14 November 2026

• Task: image to 3d
• Library: comfyui
• Downloads in the last 30 days: 0
• Likes: 11
• License: apache-2.0
• Based on: speridlabs/idol
• Tags: comfyui, gaussian-splatting, 3d, image-to-3d, idol, triposplat

Links

• Model: https://huggingface.co/PedroMarinhoDev/Idol-ComfyUI
• Files: https://huggingface.co/PedroMarinhoDev/Idol-ComfyUI/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download PedroMarinhoDev/Idol-ComfyUI

Then follow the usage in its model card. Nothing in this kit runs by itself.