Cozy Human Parser ATR

Say you've got a photo of a person and you want a clean mask of just their pants, or just their upper-clothes, without drawing anything by hand or fighting a text-prompted segmenter into understanding what "the shirt" means. That's the whole job of this node: point it at a person, flip on the categories you want, and it hands you back a mask of exactly those regions. No prompt, no bounding box, no click - just checkboxes for body parts and garments.

ATR is one of three flavors in the Cozy Human Parser pack (ATR, LIP, Pascal - same node, different training data), and it's the one built for fashion. It's trained on the ATR dataset - 17,000+ images focused on fashion AI - and it's the most accurate of the three, 82.29% mIoU against LIP's 59.36% and Pascal's 71.46%. If your work is outfits, garments, or "isolate this one clothing item," ATR is the one to reach for first.

How it works

Under the hood this is SCHP - Self-Correction for Human Parsing, a 2020 semantic segmentation model that CozyMantis wrapped into a ComfyUI node and patched to also run on CPU. It looks at your image once and labels every single pixel with one of its trained categories: hair, face, upper-clothes, and so on. That's the "map" - a full parsing of the person, every pixel classified. The node then takes whichever categories you switched on and unions those pixel regions into one grayscale mask. Nothing fancier than that: it's a fixed-vocabulary classifier, not a language model, so it can't understand "the jacket, not the shirt underneath" if both happen to fall under the same trained label - but for anything that maps cleanly onto its category list, it's fast, deterministic, and doesn't need a GPU to work.

This sits in the same neighborhood as BiRefNet/rembg cutouts and SAM/GroundingDINO targeted masking, but it does neither - it doesn't remove backgrounds and it doesn't take a text prompt. It's specifically "give me the pixels for this named body part or garment," which is the useful middle ground when you know exactly what you want and don't want to grade a VLM's guess.

Inputs and outputs

One required image input, plus a boolean toggle (default off) for each of ATR's eighteen categories: background, hat, hair, sunglasses, upper_clothes, skirt, pants, dress, belt, left_shoe, right_shoe, face, left_leg, right_leg, left_arm, right_arm, bag, scarf. Flip on whichever ones you want combined into the output mask - leave the rest off.

Two outputs: mask (MASK) - the combined mask of your selected categories, ready to feed straight into an inpainting node's mask input, a compositing node, or anywhere else ComfyUI expects a MASK. And map (IMAGE) - the full color-coded parsing visualization, useful for seeing what the model actually detected before you commit to a set of toggles.

Installing it

ComfyUI Manager: search "Cozy Human Parser," install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cozymantis/human-parser-comfyui-node
cd human-parser-comfyui-node
pip install -r requirements.txt

Then the part Manager won't do for you: download the ATR checkpoint from the Google Drive link in the README and drop it in ComfyUI/models/schp/. Restart ComfyUI once the file's in place.

Where people get burned

The install step people miss is that last one - no model file in models/schp, no working node, and the failure isn't always an obvious error. If ATR loads but every mask comes back empty, check the model actually landed in the right folder first.

The other real pain point is Windows-only, and it's a build-tools problem, not this pack's fault: pip install -r requirements.txt compiles a C++ extension, and if you're missing Visual Studio's C++ build tools, ninja, or the right python3xx.lib, you'll hit cryptic linker errors (cannot open file 'python311.lib', Command '['where', 'cl']' returned non-zero exit status). The README has a whole troubleshooting section for this - short version: install the Visual Studio C++ workload, run the install from an "x64 Native Tools Command Prompt" instead of a regular one, and make sure ninja.exe is actually on your PATH. Linux and Mac users essentially never hit this.

Worth knowing going in: this pack has almost no community footprint - its 2024 launch post got a single upvote and zero comments, and it's stayed quiet since. That's not a red flag, it's just a small, single-purpose tool that people who need it use quietly rather than something you'll see discussed on r/comfyui. If you specifically want fashion-segmentation and don't mind a newer, heavier alternative, FASHN's SegFormer-based human parser has been circulating more recently - but Cozy Human Parser stays lightweight, CPU-friendly, and needs no extra model downloads beyond the one checkpoint.