Human Garment AnimateDiff Generation

Take the MagicClothing trick - a garment photo conditioning an SD 1.5 generation - and set the whole thing in motion. That's this node. It's the same "synthesize a model wearing your clothes" pipeline as the Generate node, but with an AnimateDiff motion module bolted on, so instead of one image you get a short clip of the model turning, swaying, moving in the garment. It's the demo reel node of the pack: there is no practical reason to own it, and it's a lot of fun anyway.

Be honest about what this is, though. AnimateDiff is 2023 tech and, by any current measure, superseded - the community moved to Wan and the rest of the modern video stack years ago, and the KB's verdict is blunt: don't start a video project on AnimateDiff. What survives is a look, and this node is a very narrow flavor of it: virtual try-on, animated, at SD 1.5 quality, with all of AnimateDiff's old limitations (short clips, temporal wobble, VRAM appetite). If your end goal is "shirt on a moving person," a modern pipeline beats this in a weekend. If your goal is "weirdly satisfying 2-second cloth demo," you're in the right place.

How it works

Under the hood it builds an OmsAnimateDiffusionPipeline - a diffusers text-to-video pipeline using the guoyww/animatediff-motion-adapter-v1-5-2 MotionAdapter injected into the SD 1.5 UNet, plus MagicClothing's cloth-guidance attention adapter so the garment stays recognizable across frames. Each run produces a batch of frames (16 per clip, AnimateDiff's native context). The output is a single images tensor - not a file. To get a video you wire it into a video-save node like Save Animated WEBP or VHS VideoCombine.

The inputs that matter

Required: cloth_image, prompt, model_path (the adapter in the node's checkpoints/), pipe_path (base SD 1.5 model, Realistic Vision V4.0 by default), and motion_adapter_path (already set to the guoyww adapter - you generally leave it). The optional dials mirror the Generate node: num_images_per_prompt (how many clips per run), seed, sample_steps (20), guidance_scale (2.5), cloth_guidance_scale (3), height/width (768×576). No FaceID or pose modes here - it's the plain garment-only workflow, animated.

Installing it

Same story as the rest of the pack - 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

Then restart and drop the models in. The important difference: this node needs two extra files - garment_extractor.safetensors and ip_layer.pth - placed in checkpoints/stable_ckpt/ inside the node folder, alongside the usual cloth_segm.pth and adapter from ShineChen1024/MagicClothing on HuggingFace. The README calls this out, and missing them means the node fails at the first load - there's no graceful "download me" fallback.

Common issues

• Missing stable_ckpt files is the #1 failure mode - the code loads them by exact path, so a typo in the folder name breaks it.
• The dependency pin trap applies twice here. The pack's requirements.txt pins a 2024 CUDA 11.8 stack (torch 2.1.1, numpy 1.25.1, transformers 4.31.0), and on top of that you're running the whole AnimateDiff stack on it. A modern ComfyUI venv plus this requirements file is a conflict waiting to happen; install it with your eyes open.
• VRAM. The cloth adapter deep-copies the UNet for its reference pass, and now you're adding a motion module and 16 frames of latents on top. This is the heaviest node in the pack. Expect it to breathe on 8GB and to want more.
• First run downloads the base model, VAE, and the motion adapter from HuggingFace - several GB and an internet connection, like the other nodes.

It's the pack's least-used node (the search stats say so - it has effectively zero traffic), and it's never going to be anyone's production pipeline. But as a "here's your jacket, now it moves" party trick, it works, and it's the one to try after you've got Generate running. Just keep your expectations set to 2024.