OpenPose ControlNet guide
OpenPose is the ControlNet preprocessor for human pose. It reduces a reference figure to a skeleton of keypoints — joints connected by colored limbs — and forces the generated person to adopt that exact posture. It controls position, not appearance: the prompt still decides who the person is, what they wear, and where they are. The two decisions that matter most are which model variant you use and how much control strength you apply.
How it works
The OpenPose preprocessor detects keypoints (shoulders, elbows, wrists, hips, knees, ankles, and optionally face landmarks and finger joints) and draws them as a stick-figure map. ControlNet feeds that map alongside your prompt, biasing the diffusion process toward a person in that pose. The body-only model handles posture; the body + hands + face model adds expression and finger control (less reliably); the hands-only model targets gestures alone. Control strength scales how strongly the map constrains the result.
Model variant comparison
| Variant | Keypoints tracked | Best for | Reliability |
|---|---|---|---|
| Body only | 18 body joints | Full-figure poses, action shots | High |
| Body + hands + face | 18 body + 21 per hand + 68 face | Portraits, expressive gesture | Medium (hands error-prone) |
| Hands only | 21 per hand | Fixing gestures in otherwise good outputs | Low-medium |
Choosing the right control strength
Control strength is a trade-off between fidelity and creativity:
- 0.8–1.0 — the generated figure mirrors the reference closely. Good for product photography or reference recreation where the exact pose is required.
- 0.5–0.7 — the pose is a strong suggestion but the model adjusts proportions and angles to fit the prompt and composition naturally.
- Below 0.4 — the reference becomes a loose influence that the model often ignores. Useful mainly when you want a vague compositional hint rather than a fixed pose.
For multi-figure scenes, try a slightly lower strength (around 0.7) to give each figure room to position without clashing.
Workflow: body first, then inpaint hands
The most reliable workflow for hand-heavy compositions is:
- Generate the full figure using body-only OpenPose at strength 0.8–1.0.
- Mask just the hands and run a targeted inpaint with a detailed hand-focus prompt, higher sampling steps, and a low denoising strength (0.4–0.6).
- If an expression is critical, use body + face model instead of adding hands — face detection is considerably more reliable than finger detection.
Tips for reliable poses
- Pick the smallest model that covers your need. Body-only is the most robust; only add hands and face when you genuinely need them.
- Use strength 0.8–1.0 for faithful reproduction, 0.5–0.7 when you want the model to adapt the pose to the prompt naturally.
- Don’t trust hand keypoints. Fix the body with OpenPose, then repair hands with inpainting — detection there is unreliable.
- Clean reference, clean pose. A clear, well-lit single figure produces a far better keypoint map than a crowded or low-contrast image.
- Match image resolution. If your reference is 512×768 and your generation target is 768×512, the pose map will be cropped differently than you expect. Match the aspect ratio or use a pose editor to build the map at the right dimensions.