PIE4K pose networks (GGUF)

GGUF conversions of the person detectors and whole-body pose networks used by PIE4K's PoseService (libpie4k_posenet, a ggml runtime). They were converted from the ONNX files PIE4K previously ran with ONNX Runtime. These files are for PIE4K's own ggml graph loader and are not llama.cpp models.

Files

File	Network	Source file	Input (NCHW)	Weights	Size (bytes)	sha256
`yolox_l.gguf`	YOLOX-L person detector	`yzd-v/DWPose` `yolox_l.onnx`	1,3,640,640	as in ONNX	216730240	`dd9ca8bcf32aa472300e1d9b90e291cc42e011c1e396bf2c0ac3b135bb83f07b`
`dw-ll_ucoco_384.gguf`	DWPose-L whole-body pose (133 keypoints)	`yzd-v/DWPose` `dw-ll_ucoco_384.onnx`	1,3,384,288	as in ONNX	134387648	`4721e34ae978b4dbbd79d9f42d4c7447047de2b48bb8ad254afebdcf1f90e828`
`yolov10m.gguf`	YOLOv10-M detector	`Wan-AI/Wan2.2-Animate-14B` `process_checkpoint/det/yolov10m.onnx`	1,3,640,640	as in ONNX	61569216	`1f200913080cc2761fdd23529c94479b54d0cfda44c48eb85b5b5fc9d0a16976`
`vitpose_h_wholebody.gguf`	ViTPose-H whole-body pose (133 keypoints)	`Wan-AI/Wan2.2-Animate-14B` `process_checkpoint/pose2d/vitpose_h_wholebody.onnx/end2end.onnx` (+ external tensors)	1,3,256,192	MatMul weights fp16	1290732352	`43c62a293935f92989655382ea27017eef33a023d48d9b20cb9769f5e1411dc8`

yolox_l.gguf with dw-ll_ucoco_384.gguf is the default (DWPose) pair; yolov10m.gguf with vitpose_h_wholebody.gguf is the ViTPose pair.

Sources (pinned revisions)

yzd-v/DWPose, revision 1a7144101628d69ee7a3768d1ee3a094070dc388
Wan-AI/Wan2.2-Animate-14B, revision cb93a225fbaf1ca100f54e79da8f994995b689b3

How they were converted

tools/pose/setup_pose.sh all in the PIE4K repository downloads the ONNX files above (checked by sha256) and converts each one with tools/pose/convert_pose_onnx.py:

convert_pose_onnx.py yolox_l.onnx yolox_l.gguf --input-shape 1,3,640,640
convert_pose_onnx.py dw-ll_ucoco_384.onnx dw-ll_ucoco_384.gguf --input-shape 1,3,384,288
convert_pose_onnx.py yolov10m.onnx yolov10m.gguf --input-shape 1,3,640,640 --cut /model.23/Concat_5_output_0 --tail yolov10
convert_pose_onnx.py end2end.onnx vitpose_h_wholebody.gguf --input-shape 1,3,256,192 --f16 --name vitpose_h_wholebody

Where PIE4K looks for them

<models>/pose/, where <models> is $PIE4K_HOME/models or ~/.local/share/pie4k/models. The networks are selected in pie4k.cfg:

[pose]
PoseModel = "pose/dw-ll_ucoco_384.gguf"     # or "pose/vitpose_h_wholebody.gguf"
PoseDetectorModel = "pose/yolox_l.gguf"     # or "pose/yolov10m.gguf"