ComfyUI OpenPose ControlNet: Control Character Poses in AI Images

What is OpenPose ControlNet?

OpenPose ControlNet detects human body keypoints — head, shoulders, elbows, wrists, hips, knees, ankles — and creates a skeleton diagram. The AI then generates a new image where characters match the exact same pose.

This is the go-to ControlNet for:

• Character art — generate characters in specific action poses
• Anime and game art — consistent character poses across multiple images
• Fashion — model clothing on a body in a particular stance
• Reference matching — reproduce a pose from a photo in a completely different style

Prerequisites

Plugin Installation

You need the ComfyUI ControlNet Auxiliary Preprocessors plugin to extract poses from photos. Install via ComfyUI Manager or clone manually.

Models

File Placement

ComfyUI/
├── models/
│   ├── checkpoints/
│   │   └── dreamshaper_8.safetensors
│   └── controlnet/
│       └── control_v11p_sd15_openpose.pth

Building the Workflow

Option A: With Preprocessor (Extract Pose from Photo)

Use this when you have a regular photo and want to extract the pose from it.

1. Load Image — load a photo of a person
2. OpenPose Pose (or DWPose Estimator) — extracts the skeleton
3. Preview Image — verify the skeleton was detected correctly
4. Load ControlNet Model — loads control_v11p_sd15_openpose.pth
5. Apply ControlNet — injects pose data into conditioning
6. CLIP Text Encode (x2) — describe what you want generated
7. KSampler → VAE Decode → Save Image

Option B: With Pre-Made Skeleton Image

Use this when you already have an OpenPose skeleton image.

1. Load Image — load the skeleton image directly
2. Skip the preprocessor — connect straight to Apply ControlNet
3. Continue with the same nodes as above

Preprocessor Parameters

OpenPose Pose Node

DWPose Estimator Node (Alternative)

The DWPose estimator often produces more accurate results than the classic OpenPose detector, especially for complex poses or partially occluded bodies. It's available in the same Auxiliary Preprocessors plugin.

Key Generation Parameters

Tips for Better Poses

Clear reference photos — Use images where the full body is visible and not heavily occluded. Cropped or overlapping bodies confuse the detector.

Enable all detection — Turn on detect_hand, detect_face, and detect_body for the most complete skeleton. Disable hand/face detection only if they're causing issues.

Preview first — Always check the skeleton preview before generating. If keypoints are in the wrong position, the output will have the same errors.

Combine with other ControlNets — Pair OpenPose with Depth ControlNet for both pose control and spatial awareness. Chain the Apply ControlNet nodes in series.

Common Issues and Fixes

Pose not detected correctly

• Ensure the person is clearly visible in the reference image
• Try DWPose Estimator instead of OpenPose Pose — it handles difficult angles better
• Increase the resolution parameter on the preprocessor

Hands are deformed

• Enable detect_hand in the preprocessor
• Add deformed hands, extra fingers to negative prompts
• Lower cfg to 6–7 to give the model more freedom with hand details

Output ignores the pose

• Verify the ControlNet model file is correct (control_v11p_sd15_openpose.pth)
• Increase strength to 1.0
• Check that Apply ControlNet is correctly connected to the KSampler's conditioning inputs

Multiple people detected incorrectly

• Crop the reference image to focus on a single person
• Or embrace it — OpenPose can handle multi-person scenes if the subjects are clearly separated

SD1.5 limb errors despite using OpenPose

• OpenPose significantly reduces but doesn't eliminate SD1.5 anatomy issues
• Add comprehensive negative prompts: bad anatomy, extra limbs, missing fingers, deformed
• Consider using a checkpoint that was fine-tuned for character generation

Related Guides

• ControlNet Overview — All ControlNet types explained
• Canny ControlNet — Edge-based structure control
• Depth ControlNet — Spatial depth control