Introduction

Idol, Sperid Labs' open image-to-3D splat model, now runs inside ComfyUI and LichtFeld Studio. A new Apache-2.0 custom node pack, ComfyUI-Idol, went up on October 30 and turns one image into a native ComfyUI SPLAT, so the core RenderSplat, SplatToFile3D and SplatToMesh nodes take it from there. A day earlier a separate MIT-licensed LichtFeld Studio plugin landed that drops Idol splats straight into a LichtFeld scene with a move-rotate-scale gizmo.

We covered Idol itself when the weights dropped. These two are the first community wrappers, and they move it from a Python script into the tools prop artists and splat editors already have open.

What shipped

  • Tool: ComfyUI-Idol 0.1.0 (custom nodes) and the Idol plugin for LichtFeld Studio
  • Builders: community developers Brie-Wensleydale (ComfyUI) and lyehe (LichtFeld), wrapping speridlabs/idol
  • Modality: single image to 3D Gaussian splat, 32,768 to 1,048,576 Gaussians
  • Open vs API: open code (Apache-2.0 node pack, MIT plugin) on Apache-2.0 Idol weights, run locally
  • Where to run it today: your own ComfyUI install or LichtFeld Studio, on an NVIDIA GPU with about 8 GB free
  • Concrete fact: the node author measured 27 to 31 seconds per prop at 262,144 Gaussians on an RTX 4090, with a flat 7.2 GB VRAM peak at every Gaussian count

The ComfyUI nodes

The pack adds four nodes. Idol Loader caches the pipeline (about 3.7 GB resident). Idol Image to Splat takes an image and an optional mask and outputs a SPLAT, the 1024-pixel preprocessed image, and a path to a standard 3DGS .ply. Idol Preprocess shows the crop the model will see, and Idol Unload VRAM frees the memory when you're done.

The wiring is short:

LoadImage → Idol Image to Splat → RenderSplat (frames=60) → video
                    └→ SplatToFile3D → .ply / .splat

Two example graphs ship in example_workflows/: a five-node single-frame render and a turntable version with 24 frames. Weights (about 6.8 GB) download on the first run of the loader. The DINOv3 encoder comes from an ungated timm mirror and the FLUX.2 VAE from Comfy-Org, so you don't have to request access to any gated Hugging Face repos.

Install, per the README:

cd ComfyUI/custom_nodes
git clone https://github.com/Brie-Wensleydale/ComfyUI-Idol
python -m pip install --no-deps tensordict plyfile

Keep the --no-deps flag. The README warns that a plain install can swap out the torch in a portable build.

Settings that matter

  • mode: use synthetic for renders and generated images, and real for photos, which get a looser crop.
  • num_gaussians: from 32,768 to 1,048,576. Going to 1M only adds about 4 seconds and produces a roughly 71 MB .ply, versus about 18 MB at 262k.
  • steps 50, cfg_strength 3.0, rescale_t 3.0 are the defaults. They match upstream's presets.
  • Plug in a mask if you already have clean alpha, which skips the BiRefNet background-removal step.

The author's timing is slower than the 19.1 seconds on Idol's model card. That's because the node runs the released code path without torch.compile, and the README explains why.

The LichtFeld Studio plugin

The LichtFeld plugin is the faster route if you're laying out a splat scene rather than building a graph. You pick an image, choose Synthetic or Real photo, set a detail level from 66k to 1M Gaussians and hit Generate. Each result lands under an Idol group, and Re-decode at this detail reuses the sampled latent, so changing the Gaussian count only costs the decode step. It needs an Ampere-or-newer NVIDIA GPU. Unlike the ComfyUI pack, it asks you to accept the gated DINOv3 and FLUX.2-dev terms on Hugging Face.

Limits to know before you build a prop pipeline

  • It takes a single image only, so the model invents every unseen side. Keep turntable orbits within about 60 degrees of the input view.
  • Splats carry diffuse colour only, so metal and glass read flat.
  • You get a splat, not a game mesh. There are no UVs or rig, so run SplatToMesh if you need geometry.
  • Thin parts like antennae and wires break down.
  • The encoders keep their own licences. The FLUX.2 VAE is non-commercial and DINOv3 is under Meta's licence, so local use is fine but don't redistribute a bundle that includes them.

Why it matters for game devs

Image-to-splat props are useful for blockouts, background set dressing and quick previz. Until now Idol meant a separate script and a separate viewer. With a native SPLAT output, one ComfyUI graph can generate concept art, turn it into a splat, render a turntable and export, without leaving the canvas.

—Titus