DensePose Estimator

Where OpenPose gives you a stick figure, DensePose gives you a full body-surface map - every pixel of a person colored by where it sits on a 3D body model. The output looks like a person painted in smooth gradients of color, and it encodes far more than joint positions: it captures the orientation of the body surface itself, so it knows which way a torso is turned, how a limb is foreshortened, how the shape wraps. That extra information makes it the pose representation of choice for clothing detail, skin, and especially character animation pipelines, where a stick figure just isn't enough to keep a body coherent frame to frame.

How it works, and how it differs from OpenPose

OpenPose detects keypoints - joints, connected by lines. DensePose instead maps every visible body pixel to UV coordinates on the SMPL body model, effectively projecting a 3D body's surface onto your 2D image. The result is dense (hence the name) rather than sparse: instead of "here are 18 dots," it's "here is the entire body surface and how it's oriented." Feed that into a DensePose ControlNet and the model generates a person whose body shape and surface match, not just whose joints line up. This is why it became a backbone of tools like MagicAnimate - dense body maps drive much steadier character motion than skeletons do.

The inputs that matter

• model (default densepose_r50_fpn_dl.torchscript) - the detector backbone. The r50 default is faster; r101 is a heavier, potentially more accurate network. Start with r50; only step up to r101 if detection is missing bodies you need.
• cmap (default Viridis (MagicAnimate)) - and this one is not cosmetic. It's the color scheme of the output map, and it must match the ControlNet you're feeding. Viridis (MagicAnimate) matches MagicAnimate-style models; Parula (CivitAI) matches the DensePose ControlNet distributed on CivitAI. Pick the wrong colormap and the ControlNet reads garbage - this is the single most common DensePose mistake.
• resolution (default 512) - working size; match your render.

The single output is an IMAGE (the dense body map) that wires into a ControlNet Apply node with a DensePose ControlNet. The node makes the map; the ControlNet model is separate.

Installing it

ComfyUI Manager: search ComfyUI's ControlNet Auxiliary Preprocessors, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Fannovel16/comfyui_controlnet_aux
pip install -r requirements.txt

Restart ComfyUI. The DensePose torchscript weight downloads from HuggingFace on first run.

Where people get burned

The big one, worth repeating: cmap must match your ControlNet. If your DensePose result looks completely wrong despite a clean-looking body map, you almost certainly picked the colormap the ControlNet wasn't trained on - flip between Viridis and Parula. Second, DensePose is heavier than OpenPose and, historically, pose estimators in this pack could fall back to slow CPU execution; if it's crawling, that's the likely cause, and it's most painful on long video batches. Third, pick the right tool for the job: if you only need a person to strike a pose, OpenPose is lighter and simpler. Reach for DensePose when you specifically need body-surface fidelity - clothing that drapes correctly, consistent body shape across animation frames, the things a stick figure can't express. And, as ever, it's a hint image: no DensePose ControlNet loaded means nothing happens.