Image to pose, without the tracing
Copying a pose from a photo usually means tracing it, and a tracing is stuck at the angle the photographer chose. Photo to Pose works the other way round. It reads the position of the head, shoulders, elbows, wrists, hips, knees and ankles, estimates how each limb points in depth, and sets the same angles on a 3D mannequin. The result is a real pose you can orbit around, not a flat outline.
That makes it useful long after the import. You can view the pose from above to check foreshortening, mirror it for a facing character, swap in a wider lens for a more dramatic shot, or nudge a hand that was hidden behind the body. Artists use it to rebuild a reference they love from a new angle; animators use it to block key poses from video frames.
OpenPose skeletons for ControlNet
If you generate images with Stable Diffusion, the OpenPose output is ready to use. The PNG follows the standard COCO-18 keypoint layout and colour scheme that ControlNet's OpenPose models were trained on, drawn on black at the same size as your photo. Load it into ControlNet with the preprocessor set to none, and the generated figure takes the pose from your picture.
The OpenPose JSON uses the familiar pose_keypoints_2d format, so it opens in OpenPose editors and ComfyUI nodes that accept keypoints. Because the skeleton comes from a 3D figure, you can also turn the camera in the editor and export a new OpenPose image for the same pose from any angle, something a 2D pose estimate alone cannot do.
Tips for a clean detection
- Keep the whole body in frame, from head to feet, with a little space around it.
- One clear subject works best. With several people, the tool fits the most visible one and tells you how many it found; crop the photo to pick someone else.
- Even light and fitted clothes help. Baggy coats, heavy shadows and props in front of the body hide the joints.
- In side views the far arm and leg are partly hidden. Those parts are marked for checking, and a quick look in the editor fixes them.
- If a limb looks mirrored front to back, press its Flip button. A single photo is genuinely ambiguous about depth.
What it can and can't do
Photo to Pose fits one person per photo, the body and the general hand shape (open, relaxed, fist, pointing and similar). It doesn't copy faces, expressions or individual finger joints, and it doesn't keep anything about the person's appearance: only joint angles reach the mannequin. Photos and realistic drawings work well; heavily stylised art with unusual proportions is hit and miss.