thedeoxen/Krea-2-pose-controlnet

Krea2 Turbo – OpenPose ControlNet LoRA

📝 Short description

A classic OpenPose ControlNet LoRA for Krea2 Turbo.
Feed an OpenPose / DWPose skeleton map as the control image — the model follows the body pose and structure while the text prompt defines appearance, style, clothing, and scene.

---

📊 Examples

Each preview is a single combined image from ComfyUI: Pose → Results (left to right).

Pose → Results

---

📖 Extended description

This is a straightforward pose ControlNet-style LoRA — not a reference + pose fusion model.
It was primarily trained on humans, but also works with stylized characters.

The LoRA teaches the model to treat the input image as an OpenPose control map and to align the generated subject with that skeleton. Appearance, outfit, lighting, and environment come from the prompt (and optional style cues), not from a second reference image.

---

⚙️ How to use

1. Extract or load an OpenPose / DWPose skeleton map (black background works best).
2. Pass the pose map as image 1 in the Krea2 Ostris Edit text encoder.
3. Write a normal text prompt describing the character / scene you want.
4. Load the LoRA on the Krea2 Turbo UNet (recommended weight 0.8–1.0).

ComfyUI requirements

Pose extraction in the included workflow uses DWPose (via comfyui_controlnet_aux):

https://github.com/Fannovel16/comfyui_controlnet_aux

Also requires the Krea2 Ostris Edit nodes for conditioning with a control image:

https://github.com/ostris/comfyui-krea2-ostris-edit

Base model

Trained on and recommended with krea/Krea-2-Turbo (krea2_turbo_fp8_scaled.safetensors or equivalent).

Typical sampler settings from the example workflow: ~10 steps, CFG ~1.0, euler + simple.

✅ Example prompt

A high-resolution render of a futuristic cyber-ninja in a low-profile stealth pose. The suit is matte black with subtle blue and red LED accents. It holds two glowing plasma katanas low to the ground. The character is balanced on a high-rise ledge overlooking a rain-slicked cyberpunk megacity at night. Style: Cyberpunk aesthetics, realistic rendering, dramatic lighting.

No special trigger phrase is required — the pose comes from the control image.

---

🎯 What it does

• Follows body pose and keypoints from an OpenPose / DWPose map.
• Lets the prompt freely define identity, clothing, style, and background.
• Works as a drop-in classic ControlNet pose adapter for Krea2 Turbo.

---

⚡ Tips

• DWPose skeleton maps on a black background work best as control input.
• Keep subject scale and framing in the pose map reasonable for the scene you want.
• Lower LoRA weight (~0.6–0.8) if pose is too rigid; raise toward 1.0 for stronger adherence.
• Combine with detailed prompts for lighting, camera angle, and style.

---

📌 Use cases

• Posing characters for illustrations, comics, or storyboards.
• Enforcing a specific body pose while freely changing outfit / style via prompt.
• Animation keyframes and pose-driven concept art.
• Stylized character generation with controllable anatomy.

---

📦 Files

• Weights: krea2_turbo_openpose_controlnet.safetensors
• ComfyUI workflow: krea2_controlnet_pose.json
• Repo: thedeoxen/Krea-2-pose-controlnet