Human Garment Generation

The "only cloth image" workflow was the first thing this pack shipped, and it's still the purest version of the idea: give it a flat-lay or product photo of a garment, and it invents an entire model wearing it. You don't feed it a person at all. No pose reference, no face, no base image - one shirt in, a fully dressed person out. For a lot of people that's the whole reason they installed the pack, and it's a genuinely fun trick the first time you run it.

The catch is the name. "Generate" is the swiss-army node of the family: it also does IP-Adapter FaceID try-ons (a specific person wearing your garment) and, with a pose image added, a controlnet-openpose version that locks the body position. The basic garment-only mode is one thing; the FaceID modes are a whole other install.

How it works

Same engine as the inpainting node, minus the person. A base SD 1.5 diffusers pipeline (Realistic Vision V4.0 by default) gets a ClothAdapter bolted on: the garment is segmented (via cloth_segm.pth), encoded through the VAE, and pushed through a reference UNet whose cross-attention maps are stored and replayed during denoising. That's the "cloth guidance" that makes the fabric and print actually stick, and enable_cloth_guidance toggles it. Leave it on; turning it off just gives you a plain text-to-image with the garment as a vague hint.

The inputs that matter

Required: cloth_image, prompt (default "a photography of a model"), model_path (the adapter .safetensors in the node's checkpoints/ folder), pipe_path, and enable_cloth_guidance. Then the optional ones you'll actually fiddle with:

• num_samples, seed, sample_steps (20), height/width (768×576) - the usual sampling dials. Note the guidance knob here is called scale (default 3), not guidance_scale; the garment's own grip is cloth_guidance_scale (default 3). Two scales, two jobs.
• face_image + faceid_version (FaceID / FaceIDPlus / FaceIDPlusV2) - hand it a portrait and the model's face becomes that person. This is IP-Adapter FaceID, which swaps CLIP embeddings for InsightFace face-recognition vectors; it works, but it needs extra weights and an extra node pack, and its licensing is research-only, so don't build a storefront on it.
• pose_image - add this and it loads a control_v11p_sd15_openpose ControlNet so the person follows a pose skeleton. OpenPose conditions on structure, IP-Adapter on appearance; combining them is the classic one-two from the ControlNet playbook.

It returns two outputs: images, and cloth_mask_image - the mask of the generated garment region. Save it or reuse it as input elsewhere; it's a nice bonus you don't get from the other nodes.

Installing it

ComfyUI Manager (search "ComfyUI_MagicClothing") or:

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

Restart ComfyUI. Drop cloth_segm.pth and your chosen magic_clothing_*.safetensors (upper-body, or the lower/full-body OMS_1024_VTHD+DressCode_200000.safetensors) into the node's checkpoints/ folder. If you want the FaceID modes you also need ComfyUI_IPAdapter_plus, plus the IP-Adapter FaceID bins in models/ipadapter and their LoRAs in models/loras; the openpose mode wants comfyui_controlnet_aux and the lllyasviel annotator models. That's a lot of moving parts for an optional branch.

Where people get burned

• The dependency pin trap, again. requirements.txt wants torch==2.1.1+cu118, numpy==1.25.1, transformers==4.31.0 - a 2024 CUDA 11.8 stack. Blindly pip-installing it into a modern ComfyUI can downgrade torch and break other packs. This is the ecosystem's oldest problem and this pack is a textbook case.
• "face detection error, plz try another portrait!" - that's a real error string from the source. FaceID runs InsightFace face detection internally; a portrait with no detected face raises it. Try a clearer, front-facing photo.
• The lower/full-body model is experimental. The README says so in almost so many words: "just for experiment now," and tells you to play with hyperparameters. Expect jank.
• First run is heavy - base model, VAE, segmentation weights, and for FaceIDPlus the laion/CLIP-ViT-H-14 image encoder, all downloaded from HuggingFace. Budget the bandwidth.
• Empty model_path dropdown means the adapter isn't sitting directly in checkpoints/.

It's the node to reach for when you want a fast fashion-on-a-model mockup and don't care about a specific person. When you need the actual person, that's the inpainting node's job.