Cozy Human Parser LIP

LIP is the "everything" option of the three Cozy Human Parser flavors. Twenty categories, more than either ATR or Pascal, including a few nobody else in the pack covers - gloves, socks, coats, jumpsuits. If your subject is dressed in something that doesn't fit neatly into "upper-clothes vs pants" (someone in a full jumpsuit, someone wearing gloves you specifically want to mask), LIP is the one built to see it.

That breadth comes from the dataset it's trained on: LIP (Look Into Person) is the largest single-person parsing dataset around, 50,000+ images, and it deliberately covers "complicated real scenarios" rather than clean fashion-catalog shots. That's the trade worth understanding before you pick this one over ATR: LIP's mIoU on its own validation set is 59.36%, noticeably lower than ATR's 82.29% and Pascal's 71.46%. That's not LIP being a worse model - twenty categories across messy real-world photos is a genuinely harder task than eighteen categories on curated fashion images - but it does mean you should expect noisier mask edges out of LIP than out of ATR on the same photo. If your source images are clean studio-style shots of an outfit, ATR will usually give you a tighter mask. If your images are messy, cluttered, real-world photos and you need categories ATR doesn't have, LIP is the one that can actually see them.

How it works

Same node, same SCHP model architecture as ATR and Pascal - CozyMantis packaged Self-Correction for Human Parsing to run on CPU as well as CUDA, and each of the three nodes just loads a different trained checkpoint. It classifies every pixel of your input image into one of its trained categories in a single pass, builds a full color-coded parsing map from that, and then the node unions whichever categories you've switched on into one output mask. No prompting, no clicking regions by hand - you get a deterministic mask straight from the toggles you set.

Inputs and outputs

One required image input, plus a boolean toggle (default off) for each of LIP's twenty categories: background, hat, hair, glove, sunglasses, upper_clothes, dress, coat, socks, pants, jumpsuits, scarf, skirt, face, left_arm, right_arm, left_leg, right_leg, left_shoe, right_shoe. Switch on the ones you want combined into the mask.

Two outputs: mask (MASK) - the combined mask of everything you toggled on, the thing you route into an inpaint node's mask input or a compositing node. map (IMAGE) - the full parsing visualization across all twenty categories, worth a look before you commit, since with this many categories it's easy to guess wrong about which label a garment fell under.

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 download the LIP-specific checkpoint from the Google Drive link in the README and put it in ComfyUI/models/schp/. Restart ComfyUI once it's there - each of the three nodes needs its own model file, so if you're planning to use ATR or Pascal too, you'll be repeating this step for each.

Where people get burned

The missing-model-file trap is the most common one: no file in models/schp, and instead of a load error you just get an empty or blank mask, which is a confusing failure mode to debug from scratch. Check that folder first.

Windows adds its own, unrelated friction: the pip install -r requirements.txt step compiles a C++ extension, and without Visual Studio's build tools, ninja, and the right python3xx.lib on hand, that install throws linker errors instead of finishing cleanly. The README's Windows Troubleshooting section covers the fixes - installing the VS C++ workload, running the install from an "x64 Native Tools Command Prompt" rather than a normal one, and getting ninja.exe onto PATH. None of this touches Linux or Mac installs.

Worth setting expectations: this is a small, quiet pack. Its 2024 launch thread landed with a single upvote and no comments, and it's stayed that way - you won't find much troubleshooting discussion beyond the README itself if you get stuck. Given LIP's lower mIoU relative to ATR, if your use case is specifically fashion/outfit work on relatively clean photos, it's worth trying ATR side by side and comparing edges before settling on LIP - reach for LIP when you actually need its wider vocabulary (gloves, socks, coats, jumpsuits) or your source photos are the messier, real-world kind LIP was trained on.