Load Magic Clothing Pipeline&Diffusers

The half of this pack that lives inside diffusers

The same author's pack has two personalities. The first is the native-ComfyUI path (Load Magic Clothing Model + Add Magic Clothing Attention), where everything stays a real ComfyUI MODEL. This node is the second personality: it builds an actual HuggingFace diffusers pipeline for Magic Clothing, the "black box" style wrapper - only in this case it does the heavy lifting of converting your regular checkpoint first.

If you're on the fence about which half to use: the native path is generally the one people reach for, because it plays nice with IPAdapter and the rest of the graph. This diffusers path exists for people who either want the original pipeline's behavior, or who already work with diffusers-format models. Pick it deliberately, not because you found it first.

How it works

This is the "convert it yourself" loader. Its one model input is ckpt_name - a standard .safetensors checkpoint from your ComfyUI/models/checkpoints folder, chosen from a dropdown.

On first run it converts that checkpoint into a diffusers pipeline using from_single_file(), then saves the converted result into the pack's conversion/ folder. Every run after that just loads the cached conversion. That means the first generation is painfully slow while it writes a full pipeline directory, and it eats disk space - a converted SD1.5 pipeline is a few gigabytes. Subsequent loads are quick. If you ever switch checkpoints, budget the same conversion cost again.

Under the hood it's a subclass of the diffusers StableDiffusionPipeline that understands Magic Clothing's cloth-attention mechanism, so it can carry the garment reference through sampling the way the original project does.

Inputs and outputs

The three required inputs are mostly set-and-forget:

• ckpt_name - your SD1.5 checkpoint. The one that matters.
• dtype - default, float32, float16, or bfloat16. On a consumer GPU, float16 is the sane choice.
• device - default, cpu, cuda, cuda:0, cuda:1. default follows ComfyUI's model management; the explicit options are for multi-GPU rigs.

It outputs three wires: PIPELINE, AUTOENCODER (the pipeline's VAE), and SCHEDULER. None of them are the finish line. The PIPELINE goes into Diffusers Model Makeup &MC, which wants a SCHEDULER from Diffusers Scheduler Loader &MC and the AUTOENCODER too - then that feeds Load Magic Clothing Adapter &Diffusers, and finally RUN Magic Clothing Diffusers Model gives you an IMAGE. So this one node is just the start of a five-node diffusers subgraph. If that sounds like a lot of ceremony, the native path genuinely is simpler.

Install and what it needs

Install is the same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/longgui0318/comfyui-oms-diffusion

...or search "comfyui-magic-clothing" in ComfyUI Manager and restart. The catch: unlike the native path, this branch imports diffusers and transformers directly, and the repo ships no requirements.txt. A stock or portable ComfyUI install often doesn't have diffusers, and the node just won't register until you do:

pip install diffusers transformers safetensors

Also remember the Magic Clothing UNet weights themselves (from ShineChen1024/MagicClothing on HuggingFace) - those go into ComfyUI/models/unet and are needed by the adapter node later in the chain.

Common issues

The predictable failure modes here are the conversion ones: the first run taking forever, disk filling up with converted pipelines, and the occasional "nonstop errors" reports from back when the pack was new (mid-2024). If you only have .safetensors checkpoints, this node is the right one. If you already keep diffusers-format models, use its sibling, Load Magic Clothing Pipeline With Path, which skips conversion entirely. And remember it's SD1.5-only - Magic Clothing doesn't run on newer bases.