Load OOTDiffusion from Hub🤗

If you're trying OOTDiffusion for the first time, start here. LoadOOTDPipelineHub is the pack's convenience loader: it downloads the entire try-on model from Hugging Face, drops it in the HF cache, and builds the pipeline - all from a single type input. It's the node the pack's own example workflow uses, and it's the difference between "get the model yourself and figure out the folder layout" and "click, wait, go."

It's a subclass of the local loader, which is why it feels so familiar: same output, same heavy pipeline construction, just one less input. The only choice you make is which model variant to load.

How it works

The upstream repo levihsu/OOTDiffusion doesn't use a layout that DiffusionPipeline.from_pretrained can handle cleanly (subfolders everywhere), so this node sidesteps that with huggingface_hub.snapshot_download() - a pinned commit, revision d33c517dc1b0718ea1136533e3720bb08fae641b, so you get a known-good set of weights rather than whatever the maintainer pushed last. Everything lands in the standard Hugging Face cache, and from there the load is identical to the local loader: human parsing ONNX models, OpenPose, the two UNets, VAE, and a CLIP vision encoder, all in fp16 on your GPU.

One thing to expect on your first run: this is a big download - several gigabytes - followed by the same pipeline construction and C++ extension compile the local loader does. Give it a while, and make sure the console doesn't look hung when it's actually just downloading.

Inputs and output

• type - Half body or Full body. Same meaning as the local loader: hd (upper-body garments only) or dc (adds lower body and dresses). Pick half body if in doubt; it's the more mature variant.
• Output: a single pipe (MODEL) - the custom OOTDiffusion object that feeds straight into OOTDGenerate.

That's it. No path, no model files to manage, no manual folder structure.

The genuinely charming gotcha

Because everything lands in the HF cache, the node doesn't use a models/OOTDiffusion folder at all. If you also have one on disk - say you previously used the local loader - the node prints a warning: you've got the model in both places, and the cache copy is the one being used. The suggested move is to delete the local models/OOTDiffusion directory to free the disk space. It's a small courtesy, and it's the only thing resembling a config gotcha in this node.

Install

Same pack install as the other two nodes: ComfyUI Manager (search "ComfyUI OOTDiffusion"), or

cd ComfyUI/custom_nodes
git clone https://github.com/AuroBit/ComfyUI-OOTDiffusion
pip install -r ComfyUI-OOTDiffusion/requirements.txt

then restart. The heavy stuff - diffusers==0.24.0 pinned, CUDA + nvcc, the MSVC build tools on Windows - is pack-wide, so read the install notes in the OOTDGenerate article before you blame this node for a compile error.

Local or hub?

Honest rule of thumb: use the hub loader on first run and when you don't care where the weights live. Switch to the local loader once you've verified the results and want the model out of the cache, or when you need to run offline-ish (you'll still want the CLIP vision weights cached once). The hub node trades a few gigabytes of cache for zero file management, and for a 2024-era research pack whose install already has enough sharp edges, that's usually the right trade.