Nodes/comfyui-magic-clothing/Diffusers Model Makeup &MC
ComfyUI Node

Diffusers Model Makeup &MC

The Boring Node That Holds Magic Clothing's Diffusers Path Together

By longgui0318·Created 2 years ago·Updated 2 years ago· 82
Diffusers Model Makeup &MC
  • pipeline
  • scheduler
  • autoencoder
  • MAKED_PIPELINE

"Makeup" here doesn't mean cosmetics. It means "make up" - as in assemble. This node does zero generation and takes zero parameters; it's the point where Magic Clothing's diffusers-based branch finally becomes one usable thing. Understanding that is most of understanding the node.

Magic Clothing (from ShineChen1024's project) is garment-driven generation: you feed it a flat-ish photo of a piece of clothing and it generates a person wearing it, following the fabric's shape and print. longgui0318/comfyui-magic-clothing is the ComfyUI port, and it actually ships two ways to run it. The headline path is the native ComfyUI UNet one (Load Magic Clothing Model + Add Magic Clothing Attention) that wires into IPAdapter cleanly - that's the README's whole pitch. Then there's a secondary diffusers path bolted on, where a real HuggingFace-style pipeline does the work. This node is the glue of that second path. If you're only doing the native workflow, you can skip the entire diffusers branch and never look back.

What it actually does

Diffusers Model Makeup &MC takes three inputs - pipeline, scheduler, and autoencoder - and returns a single MAKED_PIPELINE. It's a straight copy of the "Diffusers Model Makeup" node from Limitex's ComfyUI-Diffusers pack (the &MC suffix is how this author marks their clones so the two packs don't fully collide). Under the hood it:

  • unwraps the pipeline tuple the loader returns,
  • moves your autoencoder onto the pipeline's device and dtype and installs it as pipeline.vae,
  • installs your chosen scheduler,
  • replaces the SD 1.5 safety checker with a lambda that always passes (yes, it's dead in the water before the run),
  • and calls enable_attention_slicing() to trade a little speed for VRAM headroom.

Then the assembled pipeline flows out. Nothing to tune - it's a fixup node, the plumbing between the loaders and the thing that actually renders.

The workflow it lives in

Upstream you have the two pipeline loaders - "Load Magic Clothing Pipeline With Path &Diffusers" for a folder you already have, "Load Magic Clothing Pipeline &Diffusers" for converting a normal checkpoint into a diffusers pipeline on first run. Both emit a PIPELINE plus an AUTOENCODER and SCHEDULER. "Diffusers Scheduler Loader &MC" lets you swap in a different scheduler from a dropdown (DDIM, Euler, DPM++, UniPC, the usual suspects). All three wires converge here.

Downstream, the MAKED_PIPELINE plugs into "Load Magic Clothing Adapter &Diffusers", which pairs it with the Magic Clothing UNet you dropped into models/unet, and "RUN Magic Clothing Adapter &Diffusers" does the actual cloth-conditioned sampling. So the real decisions - which VAE, which scheduler, which base model - all happen upstream of this node. This one just locks them in. The only genuinely interesting knob anywhere near it is the one that isn't here: pick the right VAE at the loader, because this node will happily swap in whatever you gave it, wrong match included.

Installing it

ComfyUI Manager: search "comfyui-magic-clothing" and hit install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/longgui0318/comfyui-magic-clothing

then restart ComfyUI. Models come from HuggingFace - grab magic_clothing_768_vitonhd_joint.safetensors and drop it in ComfyUI/models/unet. One real dependency gotcha: the diffusers path imports the diffusers Python package, and the repo has no requirements.txt to install it for you. If Manager doesn't pull it, pip install diffusers and restart. And remember this whole branch is SD 1.5-based - don't point it at an SDXL checkpoint and expect magic.

Where people get burned

  • "I installed it and can't find the node." It's a known first-run stumble. Search by "Magic Clothing" or look in the Diffusers category; the class names are exact strings like Diffusers Model Makeup &MC, which don't match what you'd guess.
  • Washed-out or wrong colors. Same trap as always (the KB's VAE essay covers it): this node swaps in whatever autoencoder you fed it. Feed it the right VAE for the base model, not a famous one from a different architecture.
  • Dense patterns come out mangled. The README is upfront that this port has a low success rate on busy, dense garment prints - a known quality ceiling, not your misconfiguration. Keep the garment photo flat, well-lit, and centered.
  • Installing it alongside other Magic Clothing ports can trigger node-name conflicts in Manager. Don't install two ports that register the same node names.

It's the most disposable node in the pack and also the one you can't skip if you're on the diffusers path. Wire it right once, forget it exists, and go fight the actual fight: getting the garment to survive the transfer.

CategoryDiffusers

Inputs (3)

NameTypeDefaultDescription
pipelinePIPELINE
schedulerSCHEDULER
autoencoderAUTOENCODER

Outputs (1)

NameTypeDescription
MAKED_PIPELINEMAKED_PIPELINE