ComfyUI‑Idol Adds Native 3‑D Gaussian Splat Support for Single‑Image to 3‑D Conversion

Overview

The ComfyUI‑Idol repository (GitHub: Nynxz/ComfyUI-Idol) provides a native implementation of the *Idol* pipeline – a TripoSplat‑based model that turns a single image of an object into a 3‑D Gaussian‑splat (3DGS) representation in a few seconds on a modern GPU.

Key Features

• Zero‑install workflow – works with stock ComfyUI nodes; no extra packages are required.
• Three custom nodes that extend the standard TripoSplat graph:

1. Idol Preprocess Image – crops, optionally removes background, and handles alpha masks.
2. Idol Camera – predicts the original camera pose so the input view can be re‑rendered or used for a turntable animation.
3. Idol Prune Splat – discards near‑transparent gaussians (opacity < 0.05) and tiny spikes, halving file size and cleaning the visual output.

• Full compatibility with existing TripoSplat support in ComfyUI 0.22+.
• Speed – on an RTX 5090, 50 Euler steps take ~18 s; decoding 262 144 gaussians adds 4‑8 s (no torch‑compile needed).

Installation & Usage

1. Clone the repo into `ComfyUI/custom_nodes` and restart ComfyUI.
2. Ensure your ComfyUI installation already includes TripoSplat (v0.22 or newer).
3. Load the example workflow `example_workflows/idol_image_to_splat.json` via Templates → Custom Nodes → ComfyUI‑Idol.
4. Adjust the crop mode:

• *Synthetic* (object fills ~1/1.2 of the frame) for rendered images.
• *Real* (object fills ~1/1.7) for photographs.

5. Connect an input image (or PNG with alpha) to Idol Preprocess Image. If the image lacks a mask, BiRefNet can generate one automatically.
6. Run the graph; the Idol Camera node will output a camera description that can be fed to Render Splat for a view‑matched preview.
7. Use Idol Prune Splat before saving to reduce haze and file size.
8. Export with Save Splat (PLY, SPZ, KSplat) or convert to a mesh via Extract Mesh (GLB).

Practical Tips

• Mask quality matters – a clean cutout prevents artifacts in the 3‑D view.
• Density settings – 262 144 gaussians is the default; higher values increase VRAM usage without adding detail, while 65 536 offers a fast preview.
• Transparent PNGs – pass the image’s alpha channel through an *Invert Mask* node into Idol Preprocess Image to keep original transparency.
• Verification – compare the rendered view from Idol Camera with the original image; mismatches usually indicate an upstream mask or crop‑mode issue.

Licensing

• The ComfyUI‑Idol pack is released under Apache‑2.0.
• Underlying components retain their original licenses: DINOv3 (Meta), FLUX.2 VAE (Black Forest Labs), BiRefNet (Zheng et al.), and the TripoSplat architecture (VAST‑AI‑Research).

How to Confirm the Announcement

• Visit the GitHub repository at https://github.com/Nynxz/ComfyUI-Idol
• Check the `README.md` and the `example_workflows` folder for the referenced JSON workflow.
• Review the `tests/parity.py` and `tests/e2e.py` scripts, which validate numerical parity with the reference Idol SDK.

--- *This article summarizes the publicly available repository information; no independent testing was performed.*