---
license: apache-2.0
base_model: speridlabs/idol
pipeline_tag: image-to-3d
library_name: comfyui
tags:
  - comfyui
  - gaussian-splatting
  - 3d
  - image-to-3d
  - idol
  - triposplat
---

<img alt="ComfyUI-Idol: one image to a 3D Gaussian splat, natively in ComfyUI" src="assets/banner-light.png" width="100%">

<p align="center">
  <img alt="ComfyUI v0.43+" src="https://img.shields.io/badge/ComfyUI-v0.43%2B-64B5F6?style=flat-square&labelColor=555">
  <img alt="VRAM about 6 GB" src="https://img.shields.io/badge/VRAM-~6%20GB-64B5F6?style=flat-square&labelColor=555">
  <img alt="Apache-2.0" src="https://img.shields.io/badge/license-Apache--2.0-64B5F6?style=flat-square&labelColor=555">
</p>

<p align="center">
  <a href="https://github.com/PedroMarinhoDev/ComfyUI-Idol"><b>Nodes on GitHub</b></a> ·
  <a href="https://huggingface.co/speridlabs/idol"><b>The original model</b></a> ·
  <a href="https://github.com/speridlabs/idol-sdk"><b>Idol SDK</b></a>
</p>

**[Idol](https://huggingface.co/speridlabs/idol)** by speridlabs, converted for ComfyUI's native TripoSplat nodes:
one picture of an object in, the whole object out as a 3D Gaussian splat (up to 1,048,576 Gaussians), including the sides
the picture never showed. With the **[ComfyUI-Idol](https://github.com/PedroMarinhoDev/ComfyUI-Idol)** nodes it runs
on ComfyUI's own Load Diffusion Model, KSampler and TripoSplat Decode, with ComfyUI's memory management, not a wrapper
around the SDK.

<img alt="Five objects turning: a vinyl figure, a tin robot, a treasure chest, a toadstool cottage and a bonsai, each with its input image under it" src="assets/showcase-light.webp" width="100%">

<sub>Idol, <code>idol_fp32</code> at the default precision · 262,144 Gaussians · 50 steps, cfg 4, seed 42 · RTX 3090 · each turntable starts at the input's viewpoint (Idol Input Camera) · inputs made with Krea 2 Turbo from our own prompts</sub>

## Files

| File | Folder | Size | |
|---|---|---|---|
| [`diffusion_models/idol_fp32.safetensors`](diffusion_models/idol_fp32.safetensors) | `models/diffusion_models` | 1.48 GB | the denoiser, as released; needed for the fp32 presets |
| [`vae/idol_vae_decoder_fp32.safetensors`](vae/idol_vae_decoder_fp32.safetensors) | `models/vae` | 1.15 GB | the octree and Gaussian decoders in one file, as released |
| [`diffusion_models/idol_fp16.safetensors`](diffusion_models/idol_fp16.safetensors) | `models/diffusion_models` | 0.74 GB | the denoiser in fp16: same speed and same result at the default precision |
| [`vae/idol_vae_decoder_fp16.safetensors`](vae/idol_vae_decoder_fp16.safetensors) | `models/vae` | 0.58 GB | the decoders in fp16 |

Idol also reads the image through DINOv3 ViT-H+/16 and the FLUX.2 VAE, and removes the background with BiRefNet: the
same files ComfyUI's TripoSplat template uses, from [VAST-AI/TripoSplat](https://huggingface.co/VAST-AI/TripoSplat).

## Use

1. Install **[ComfyUI-Idol](https://github.com/PedroMarinhoDev/ComfyUI-Idol)**: search for **Idol (image to 3D splat)**
   in ComfyUI-Manager, run `comfy node install comfyui-idol`, or clone it into `ComfyUI/custom_nodes` (no pip install).
   Restart ComfyUI 0.43 or newer.
2. Download the files:

   ```bash
   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 .
   ```

3. Open **Templates → ComfyUI-Idol → idol_image_to_splat**, pick your image in Load Image, run. Idol's settings are
   set: euler, simple, 50 steps, cfg 4 (Idol's guidance 3 in ComfyUI's convention).

## What changed from the release

The tensors are the release's (`denoiser.safetensors`, `decoder.safetensors`, `decoder_gs.safetensors` of
[speridlabs/idol](https://huggingface.co/speridlabs/idol)); the fp32 files hold them tensor for tensor, the fp16 files
cast to fp16.

- The denoiser's stored token positions (`pos_pe`) are left out: ComfyUI computes the same scrambled Sobol sequence,
  and the converter checks the two are equal.
- The two decoders are one file, under the `octree.` and `gs.` prefixes ComfyUI's TripoSplat decoder expects.

Every key and shape loads into ComfyUI's TripoSplat classes. One denoiser forward in fp32 matches the reference SDK to
3.5e-6 relative error, the decoders to 9e-5; the node pack's README has the full comparison and the converter
(`tools/convert_idol.py`).

## License

Apache-2.0, as the original weights: Idol by [speridlabs](https://speridlabs.com). The architecture is TripoSplat by
VAST.