UnblendPoseEngine
Pose extraction from anime art with an engine choice.
pose_engine:
dwpose — imgutils DWPose (COCO, real people). The current standard.
rtmlib_humanart — RTMPose halpe26 + HumanArt (sees anime). Recommended for anime.
Catches anime proportions better (big head, long legs, etc.).
device:
cpu (default) — safe, no onnxruntime CUDA errors.
cuda — tries to use the GPU via onnxruntime.
On CUDA 13 dev / RTX 50xx it often fails with "cublasLt64_12.dll missing".
In that case the node automatically runs on CPU and prints a
detailed warning to the console once. Quality stays high.
det_conf (rtmlib_humanart only):
The YOLOX detector threshold. Default 0.3 (the original rtmlib used 0.7).
For duo art (2+ characters) lower it to 0.25-0.4 so the second character isn't dropped.
If you use out_bboxes_json (boxes from DeepghsCharSeg / AnimeCharDetectSeg),
the detector does not run at all — poses are computed exactly from your character boxes.
This is the best way for duo/trio art.
out_bboxes_json:
Wire the boxes from your character detector here.
Then rtmlib and dwpose extract poses for exactly these regions (a "one per character" guarantee).
Outputs:
pose_image : IMAGE — OpenPose RGB (all skeletons on one image) for ControlNet
bodies_json : STRING — a JSON list of bodies (box/seed/points/conf/body18)
compatible with DeepghsCharSeg detect_target=pose
Unblend · 2 Ausgaben
