Stable-3D Load Models

Stable-3D Load Models is the unglamorous front door of the ComfyUI-Stable3DGen pack. It doesn't generate anything and it doesn't preprocess anything - it downloads the weights every other node needs, loads them onto your GPU, and hands back the two connections that run through the whole workflow. Boring, yes. It's also where first-run patience goes to die, because this node quietly pulls down gigabytes the first time you execute it.

What it actually does

Three model families land in ComfyUI's models/trellis folder (the node registers that folder and creates it for you):

• Stable-X/trellis-normal-v0-1 - the Hi3DGen generation pipeline weights
• Stable-X/yoso-normal-v1-8-1 - the YOSO-based normal-prediction model
• ZhengPeng7/BiRefNet - background removal, the same model that's become this ecosystem's default cutout and that ComfyUI itself shipped in core in 2026

Then it pulls two torch.hub dependencies into ~/.cache/torch/hub: hugoycj/StableNormal (StableNormal_turbo) and facebookresearch/dinov2 (dinov2_vitl14_reg, the DINOv2 features the normal model leans on). Finally it builds the pipeline, moves it to CUDA, and attaches BiRefNet to it. The node checks whether a folder already exists before downloading, so a failed first run doesn't force a full re-download.

The inputs and outputs that matter

Three STRING inputs, all Hugging Face repo ids with sensible defaults already filled in: trellis_model, normal_model, birefnet_model. For a beginner this is a leave-it-alone section - you'd only touch these to point at a different checkpoint or a BiRefNet variant like the HR weights. If you do swap one, know that the folder it lands in is named trellis, which is a bigger hint about the pack's DNA than any documentation: Stable3DGen is Stable-X's Hi3DGen, a fork of Microsoft TRELLIS with the NVIDIA-only libraries (kaolin, nvdiffrast, flexicube) stripped out so it's clean MIT and commercially usable. lerignoux's wrapper just vendors that whole upstream project as a git subtree.

The two outputs are the handles the rest of the graph needs:

• hi3dgen pipeline (HI3DGEN_PIPELINE) - goes into both Preprocess Image and Generate 3D
• Normal predictor (STABLE3D_NORMAL) - goes into Preprocess Image

Install

ComfyUI Manager: search "Stable3DGen" (it's published to the Comfy registry under the title "ComfyUI Stable3DGen"). Or the old-fashioned way:

cd ComfyUI/custom_nodes
git clone https://github.com/lerignoux/ComfyUI-Stable3DGen
cd ComfyUI-Stable3DGen
python -m pip install -r requirements.txt

Then restart ComfyUI. That requirements.txt deserves a read before you run it - it pins torch==2.5.1, torchvision==0.20.1, spconv-cu120==2.3.6 and xformers, which can clash with a ComfyUI install already running a different torch version. The upstream Stable3DGen project recommends a fresh Python 3.10 environment for exactly this reason. spconv (a sparse-convolution library with a CUDA 12 build) is the most common Windows failure point - you'll see it die at import or at build time, rarely with a useful message. Load Models also wraps itself in SPCONV_ALGO=native, an env-var shim for newer NVIDIA architectures.

Gotchas

First run looks like a hang. It's not - that's snapshot_download grinding through the trellis checkpoint. Let it cook once and the node is seconds-fast forever after. If you'd rather not sit through it in the UI, pre-download into ComfyUI/models/trellis yourself and it'll skip straight to loading. And note it's CUDA-only: no pipeline without an NVIDIA GPU, and this isn't a low-VRAM hobby - the generation side of this pack wants real memory, so budget accordingly.