| --- |
| license: other |
| library_name: safetensors |
| tags: |
| - comfyui |
| - pose-estimation |
| - object-detection |
| - wholebody |
| pipeline_tag: keypoint-detection |
| --- |
| |
| # BCVideoNodes models |
|
|
| The detection and pose models used by |
| [ComfyUI-BCVideoNodes](https://github.com/0xBeycan/ComfyUI-BCVideoNodes). |
|
|
| The nodes download whatever they need from here on first use, so nothing in this repository |
| has to be fetched by hand. It exists so the weights have a stable home, and so anyone can |
| see exactly which files a node is pulling. |
|
|
| ## What is here |
|
|
| | file | model | source | licence | |
| |---|---|---|---| |
| | `vitpose_h_wholebody_fp16.safetensors` | ViTPose-H, 133 COCO-WholeBody keypoints | [Kijai/vitpose_comfy](https://huggingface.co/Kijai/vitpose_comfy) | Apache-2.0 | |
| | `rtmw_l_wholebody_384x288_fp32.safetensors` | RTMW-l, 133 COCO-WholeBody keypoints | OpenMMLab's own export, via [bukuroo/RTMW-ONNX](https://huggingface.co/bukuroo/RTMW-ONNX) | Apache-2.0 | |
| | `yolov10x_fp32.safetensors` | YOLOv10-X person detection | [onnx-community/yolov10x](https://huggingface.co/onnx-community/yolov10x) | AGPL-3.0 | |
|
|
| `vitpose_h_wholebody` is fp16; the other two are fp32. See below. |
|
|
| ## Why these are not the original files |
|
|
| Upstream ships these as ONNX. The nodes run them as plain torch modules instead — the |
| architecture is written out by hand and the weights are loaded into it, which removes the |
| `onnx` dependency, removes a protobuf parse and a graph walk from every model load, and |
| replaces ViTPose's two-file external-data export (a 420 KB graph plus a 2.5 GB `.bin`) with |
| one file. |
|
|
| Each file carries its architecture in the safetensors metadata, so the loader needs nothing |
| but the file. |
|
|
| ## Precision |
|
|
| ViTPose-H is fp16, the other two are fp32, because fp16 is not free and only one of them is |
| large enough for it to pay: |
|
|
| | model | fp32 | fp16 | saved | |
| |---|---|---|---| |
| | ViTPose-H | 2.43 GB | 1.22 GB | 1.2 GB | |
| | RTMW-l | 219 MB | 110 MB | 110 MB | |
| | YOLOv10x | 113 MB | 57 MB | 56 MB | |
|
|
| What it costs, measured on 60 real frames over every keypoint the model draws, against the |
| same model in fp32 on identical crops: ViTPose-H, 5230 keypoints compared, 20 moved, largest |
| move **1 heatmap cell**. RTMW-l, 5012 compared, 401 moved, largest move **1.5 px**. Worth |
| paying for 1.2 GB, not worth paying for 110 MB. |
|
|
| Keypoints the model does not draw are excluded from those figures, and that matters: on |
| those the score distribution is flat, so the argmax is noise in either precision, and fp16 |
| appears to move them by up to 215 px. That number is what a flat distribution does, not what |
| fp16 does. |
|
|
| ## Conversion |
|
|
| The converter lives in the node repository, |
| [`scripts/convert_models.py`](https://github.com/0xBeycan/ComfyUI-BCVideoNodes/blob/main/scripts/convert_models.py). It reads the upstream ONNX file, builds the |
| native module from it, and writes the module's own state dict plus its architecture. Rerun |
| it to update a model; do not hand-edit the files here. |
|
|
| ## Credit |
|
|
| The pose models are OpenMMLab's and the ViTPose authors' work; YOLOv10 is |
| [THU-MIG](https://github.com/THU-MIG/yolov10). This repository only re-packages them. The |
| wholebody pose utilities in the node repository are vendored from the Alibaba Wan team's |
| WanAnimate preprocess and keep their copyright header. |
|
|
