Krea-2 Pose Control-LoRA
Pose-conditioned image generation for Krea-2 using a 17-keypoint body skeleton.
The pose condition controls broad body geometry and placement, while the prompt controls compatible subject appearance, clothing, environment, lighting, and style.
Model
Recommended release file:
release/krea2-pose-control-mix025.safetensors
- Candidate:
mix-025 - Saved parameters:
215,488,512 - SHA-256:
6d97e9c2e102e07928fc8864346401a0d2e6082d610ca6b037c4704102e3f8d1
Frozen inference uses Krea-2 Turbo. The control mechanism is an expanded
ControlInputLayer plus rank-64 LoRA, not a parallel ControlNet branch.
The other checkpoint folders are archival historical training runs.
release/ contains the recommended public model.
ComfyUI
For ComfyUI, use the standalone package:
https://github.com/s-adhit/ComfyUI-Krea2-Pose
It includes the custom nodes, PoseBridge preprocessing, and an example reference-image workflow.
The node pack includes:
Krea2PoseExtractor— reference image to PoseBridge COCO-17 pose conditionKrea2PoseCondition— validates an existing pose conditionKrea2PoseGenerate— pose-conditioned Krea-2 Turbo generation
Pose conditioning
The model uses 17 COCO-style body keypoints:
- nose, eyes, ears
- shoulders, elbows, wrists
- hips, knees, ankles
Hands, fingers, and face landmarks are not part of the control signal.
Use pose rasters generated in this project's trained format rather than assuming arbitrary OpenPose renders are compatible. Let the pose set body geometry and framing; keep the prompt compatible with its pose and subject count rather than narrating joints.
Inference
Frozen inference defaults:
Base: Krea-2 Turbo
Steps: 8
CFG: 0
mu: 1.15
Control scale: 1.0
Geometry: native aspect-preserving
PYTHONPATH=. python inference.py \
--turbo-ckpt /path/to/krea2-turbo.safetensors \
--release-artifact release/krea2-pose-control-mix025.safetensors \
--prompt "fantasy mage, ornate robes, cinematic lighting" \
--pose-image /path/to/pose.png \
--output output.png \
--seed 42
For stronger pose adherence, a control scale of 1.25–1.50 can also be useful.
See the prompting guide for examples and prompt recommendations.
Results
| Model | PCK@0.05 | PCK@0.10 | PCK@0.20 | CLIP |
|---|---|---|---|---|
| Turbo base | 0.0356 | 0.1101 | 0.3387 | 0.34402 |
| mix-025 | 0.4521 | 0.6043 | 0.7241 | 0.33694 |
Native aspect-preserving inference gave better pose adherence than dynamic-768 in the frozen evaluation.
Limitations
- Overlapping multi-person poses are harder.
- Fingers and facial pose are not directly controlled.
- Prompts that conflict with the supplied pose, framing, or subject count can reduce pose adherence.