(Down)load MoGe Model

This is the boring half of a two-node pack, but you can't skip it: DownloadAndLoadMoGeModel is what actually gets the MoGe weights onto your machine and onto your GPU. MoGe is Microsoft Research and Tsinghua's monocular geometry estimator - the model that takes one flat image and recovers real 3D geometry from it, as a per-pixel point map rather than just a depth map. Kijai wrapped it in exactly two nodes, and this one is the mandatory first step in every MoGe workflow.

What makes it nice (and slightly unusual) is that the download is fully automatic. The name is a lie in the good way: you don't pick a local file, you pick from a dropdown, and the node pulls it from Hugging Face on first run if it isn't already sitting in ComfyUI/models/MoGe. No manual download, no wget dance, no hunting for the right file.

How it works

The node checks models/MoGe for the file you selected. Missing? It calls snapshot_download against Kijai/MoGe_safetensors and only fetches the pattern matching your pick. Present? It loads the safetensors with ComfyUI's load_torch_file, builds the model with a hardcoded config (DINOv2 ViT-L encoder plus a convolutional decoder), and drops it on your torch device - using accelerate for the transfer when it's available.

There's nothing to configure on the model side, which is the point. The architecture is baked in; this node's whole job is "make the model exist on the GPU."

The two inputs that matter

• model - two choices: MoGe_ViT_L_fp16.safetensors or MoGe_ViT_L_fp32.safetensors. That's the whole model menu. fp16 runs ~628 MB, fp32 ~1.26 GB, and for mesh extraction from a still image the difference is in the noise. Pick fp16 unless you have a reason not to.
• precision - fp16 / fp32 / bf16, default fp32. The tooltip warns it "has no effect with GGUF models," which is a copy-paste leftover from Kijai's video-model packs: there are no GGUF files here, so ignore that line. In practice, if you loaded fp16 weights, leave precision on fp16.

Output

One output: moge_model, typed MOGEMODEL. There's exactly one thing it plugs into - the other node in this pack, MoGeProcess. That's the whole wiring diagram, which is refreshing after packs with twenty intermediate types.

Installing it

Same as every Kijai pack. Via ComfyUI Manager, search "MoGe" (the pack title is ComfyUI-MoGe) and install, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-MoGe

Restart ComfyUI. Its requirements.txt asks for trimesh, pillow, scipy, numpy, huggingface_hub, and opencv-python - all common, nothing exotic. The model itself auto-downloads to ComfyUI/models/MoGe on first run.

Troubleshooting

The main thing that bites people is the first-run download: it needs to reach Hugging Face, and it isn't instant. If you're offline or HF is down, grab the safetensors from the Kijai/MoGe_safetensors repo manually and drop it into ComfyUI/models/MoGe with the exact dropdown filename. Beyond that, this node is about as fuss-free as Kijai packs get - the real quirks live in MoGeProcess, where the actual geometry work happens.