DWPose Settings Explained: Resolution and Detection

The DWPose Estimator node looks almost empty: a detector choice, one resolution number, and three toggles. Yet those few controls decide whether your skeleton has fingers or mittens, whether faces come through, and whether a run takes two seconds or twenty. People often blame dwpose itself for a bad skeleton when the real culprit is a setting left at a default that does not suit their photo. This guide explains every control on the node in plain language, what each one actually changes inside the estimation process, and the exact combinations worth using for portraits, full-body shots and crowded scenes.

Where the Settings Live

In ComfyUI, the settings sit on the DWPose Estimator node from the comfyui_controlnet_aux pack: a bbox_detector dropdown, a detect_resolution number, and switches for body, hands and face detection. In AUTOMATIC1111, the same ideas appear in the ControlNet unit when you pick a dwpose preprocessor, with resolution exposed as a slider. The names differ slightly between the two homes, but the machinery underneath is identical, so everything in this guide applies to both. Change one setting at a time while you learn; skeletons are quick to regenerate, and single changes teach you exactly what each control does.

bbox_detector: The Person Finder

Before dwpose can measure anyone, something has to find them. The bbox detector (a YOLOX model) scans the full image and draws a bounding box around each person, and the pose model then works inside those boxes. The dropdown lets you pick which detector variant to use, or switch detection off entirely. Leaving the standard detector on is right for almost every photo, especially anything with multiple people, cluttered backgrounds or a subject who takes up only part of the frame. Turning it off makes the pose model treat the whole image as one subject, which occasionally helps for a tightly cropped single portrait, and quietly breaks everything else. If your skeletons ever show a person merged with background furniture, or a second person is missing entirely, this setting and its model are the first place to look.

detect_resolution: The Quality Dial

This is the setting you will touch most. Detect resolution controls the size at which the estimator processes your image internally. The common default is 512, which is fast and perfectly fine for large, clear subjects. Raising it makes estimation slower and noticeably sharper, and the difference concentrates exactly where dwpose earns its keep: fingers, facial landmarks, feet, and anyone standing far from the camera. A hand that occupies thirty pixels in your photo carries almost no information at low resolution; at higher resolution the estimator has real detail to work with and the finger fan appears. Lowering the value below default speeds things up for rough drafts and video previews, at the cost of wobbly small parts. Treat 512 as your starting point, 768 or higher as your “hands matter” setting, and anything lower as a speed tool.

The Detect Body Toggle

Body detection is the core skeleton: the 17 body points plus feet that form the main stick figure. You will leave this on for essentially every job, since hands and face points are positioned relative to a body that has to exist first. The rare exception is experimental work where you only want a face or hand study, and even then most people keep the body on and simply ignore the extra lines. If you ever see a skeleton with floating hands and no arms, you have found the one workflow where somebody turned this off by accident.

The Detect Hands Toggle

Hands are the most expensive and most valuable part of the estimate. With hand detection on, dwpose adds 21 points per hand and draws the finger fans that make gestures readable to ControlNet. Turning it off speeds the run a little and produces cleaner, simpler skeletons for distant full-body shots where hands are tiny smudges anyway. The rule of thumb is honest: if the hands are visible enough that a viewer could tell what they are doing, keep detection on and raise the resolution instead of giving up. If the subject is a speck in a landscape, hands off is fine and your skeleton will look less noisy.

The Detect Face Toggle

Face detection adds the 68 facial landmarks: jaw line, brows, eyes, nose and lips. For portraits and character work this is gold, because ControlNet receives the head’s tilt and turn in real detail rather than guessing from a blank circle. For wide shots it contributes little, since the face may only be a few pixels across, and the extra points can clutter the skeleton. Match the toggle to the shot. Portraits and close-ups: on, with healthy resolution. Crowd scenes and distant figures: off, and nobody will miss it.

Recipes for Common Jobs

Portrait or character close-up: body, hands and face on; detect resolution 768 or higher; standard detector on.
Full-body fashion or dance reference: everything on; resolution 640 to 768 so fingers and feet stay honest at distance.
Quick pose draft: face off, hands on; resolution at the 512 default; speed over polish while you scout references.
Group photo: detector definitely on; face and hands on only if the people are large in frame; resolution raised one step to help smaller figures.
Video frame sequences: pick one recipe and lock it for the whole clip; consistency across frames matters more than squeezing out detail. The video guide below explains why.

How Resolution Interacts with Your Photo

One subtle point catches people out: detect resolution works relative to what your photo actually contains. A giant, sharp reference photo downscaled to 512 still loses the tiny hand detail, because the estimator only ever sees the processed size. Conversely, feeding a small, blurry photo at a huge detect resolution cannot invent detail that was never captured; it just takes longer to tell you so. The winning combination is a decent source image plus a resolution matched to the smallest body part you care about. When in doubt, run the estimator twice at two resolutions and compare the previews. Ten seconds of comparison beats an hour of guessing, and the difference in the hands is usually obvious at a glance.

Settings Mistakes That Look Like Bugs

A few configuration slips produce symptoms people misread as broken software. A skeleton missing its second person is usually the detector being off or an old detector choice, not a corrupt model. Hands that appear on some runs and vanish on others usually trace back to the source image changing size, because resolution interacts with what the photo contains, as explained above. A face drawn with points but a head pointing the wrong way is the estimator struggling with a heavily turned or shadowed face; raise the resolution and improve the reference before blaming the toggle. And a run that suddenly takes five times longer almost always means somebody pushed detect resolution to its ceiling for a whole batch of large images. Every one of these is a settings conversation, not a reinstall.

Conclusion

The DWPose Estimator’s short settings list hides a simple logic: the detector finds people, the resolution decides how much detail the model gets to see, and the three toggles decide which parts of the 133-keypoint layout you want drawn. Keep the detector on, treat 512 as a starting resolution rather than a law, switch hands and face on whenever they are visible enough to matter, and lock your choices per project instead of fiddling mid-series. Master those habits and dwpose stops producing occasional skeletons and starts producing the skeleton you expected, every time.

Read next:

DWPose Settings Explained: Resolution and Detection Options