Cozy Pose Body Reference

The name is a slight lie, in the good way. Cozy Pose Body Reference doesn't detect or generate any poses - it hands you a clean, hand-drawn OpenPose-style skeleton and calls it a day. No API, no key, no model to download, no image to feed it. You pick an angle and it gives you the exact stick figure you'd get if you ran DWPose on a perfect studio photo, minus all the wobble and double-joints a real detector sometimes produces. If you've ever burned an hour fixing a broken hand skeleton extracted from a reference image, you already know why that's worth something.

What it's actually for

This is a reference-sheet tool. When you're building a character turnaround - the same character rendered from eight angles so you can train a LoRA or keep a character consistent across a sheet - the whole trick is that every panel's skeleton has to be the same figure. Extract poses from eight different photos and you'll get eight different body proportions. These PNGs are pre-authored and symmetrical, so your 8-angle sheet lines up in a way no detector-derived set ever will. That's exactly the workflow the character-consistency crowd has been doing since the multi-view/turnaround-sheet era: lock the pose, let ControlNet force the body structure, then keep identity stable with a character LoRA on top.

Honest caveat, because it matters: the skeleton only guarantees body structure. Your character still won't look identical across all eight angles without a character LoRA doing the identity work. This node is the scaffolding, not the whole building.

How it works

Open the source and there's nothing to it. The node is a file loader wearing a trench coat: it joins poses/body/<size>/<angle>.png, opens it with PIL, normalizes to float, and returns it as an IMAGE tensor. The pack ships 16 PNGs total - eight angles at 512 and 1024. Zero inference, zero heavy dependencies (just torch, numpy, pillow - all things a working ComfyUI already has). It runs in milliseconds.

The two inputs that matter

• angle - an INT from 0 to 315 in steps of 45. That's the eight canonical directions of a turnaround: front, 3/4, side, back, and the rest. There's no slider magic here; 45° is the granularity you get. If you need a 22.5° view, this node isn't it.
• size - 512 or 1024. Match it to your ControlNet's comfort zone and your VRAM. 1024 gives the pose more headroom on SDXL-era pipelines.

The output is a single pose (IMAGE). That's the whole interface.

Wiring it up

Drop the pose output into the image input of an Apply ControlNet node loaded with an openpose-capable ControlNet - the classic SD 1.5 control_v11p_sd15_openpose, an SDXL pose model, or a union model's pose mode. The key move: don't run a preprocessor on it. The image is already a pose map; piping it back through DWPose to "extract" the pose from a stick figure is redundant work that can add its own noise. Straight from this node into Apply ControlNet, then conditioning into your KSampler.

Installing it

Easiest path, ComfyUI Manager → Install Custom Nodes → search "Cozy Reference Pose Generator" (the pack title), install, restart. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/cozymantis/pose-generator-comfyui-node

The README suggests cloning into custom_nodes/cozy-pose-generator and running pip install -r requirements.txt - either folder name works since the node uses relative imports. The requirements file is just torch/numpy/pillow, so on any install that's already generating images, that pip step is basically a no-op.

Gotchas

Where people get burned: they install this expecting a pose detector and then wonder why feeding it their character photo does nothing. It's a reference generator - if you want to mirror a specific photo's pose, you want the DWPose preprocessor instead. Second gotcha: don't expect in-between angles; 45° steps are the entire vocabulary. And one small editorial note - the broader 2026 community has mostly moved character re-posing to edit models like Qwen-Image-Edit. If you're on that stack, you probably don't need this. If you're still living in SD 1.5/SDXL ControlNet land - which is where clean turnarounds still shine - this is a tidy, opinionated shortcut for a genuinely fiddly job.