RUN OMS

RUN OMS is the "press play" node from this pack's earlier, diffusers-based life - feed it a photo of a garment, get a generic model wearing it. No prompts, no conditioning graph, no fuss. If you've ever loaded a 2024-era Magic Clothing workflow, this is probably the node doing the heavy lifting in the middle of it.

A quick history so this makes sense. The pack started life as a diffusers port of the Magic Clothing project (that's the "OMS" in the repo name - the internal OmsDiffusionPipeline implementation that shipped before the native ComfyUI rewrite). RUN OMS is the run node of that era. The current README is explicit that the modern path is not the diffusers version, and the OMS nodes have since been removed from the codebase. But comfy.icu indexes every node that ever shipped, so this one lives on - and so do the workflows that use it.

What it takes and gives

• cloth_image (IMAGE) - your garment photo, cut out of its background.
• model (MODEL) - the output of LOAD OMS. Don't let the wire type fool you: this is actually a ClothAdapter object, a whole diffusers pipeline wearing a MODEL-colored cable.
• seed (default 1234), scale (CFG, default 5), cloth_guidance_scale (default 2.5), steps (default 25), height (768), width (576).

Output is a single IMAGE.

How it works

The mechanism is classic Magic Clothing. The node VAE-encodes the cloth image, runs a reference UNet on it to capture the garment's attention features, then denoises using a three-way classifier-free guidance:

uncond + scale * (text - cloth) + cloth_guidance_scale * (cloth - uncond)

The text branch is the interesting bit - it's hardcoded. The positive is "a photography of a model,best quality, high quality" and the negative is "bare, monochrome, lowres, bad anatomy, worst quality, low quality". You can't prompt your way out of it. You get a model shot, period. That's by design: the clothing comes from the image, the person comes from the base model, and the text is just there to keep the whole thing on-model.

It's also defensive: if you wire something that isn't a ClothAdapter into the model slot, the node silently returns your input image unchanged instead of crashing.

The inputs that actually matter

Two. cloth_image quality decides the garment fidelity - a clean, single-color cutout beats a busy photo every time, and this method is known to choke on dense patterns. And cloth_guidance_scale is the dial for how hard the garment sticks: too low and it drifts toward "a shirt that happens to be similar," too high and the jacket starts eating the person.

Install and models

Shared with the whole pack: ComfyUI Manager → "comfyui-magic-clothing", or:

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

LOAD OMS (the loader for this node) hardcodes Realistic_Vision_V4.0_noVAE as the base model and pulls it from HuggingFace on first run - that's a hidden multi-gigabyte download that catches people off guard - and it needs the Magic Clothing UNet file in ComfyUI/models/unet. The "model not found" error you'll see in help threads is almost always the UNet sitting in models/checkpoints instead. And keep your expectations in the pack's own stated range: low success rate, rerolls expected. When it hits, though, it's the whole virtual-try-on pipeline in one node.