jtydhr88/ComfyUI-InstantHMR

English | 中文

InstantHMR for ComfyUI

InstantHMR weights repackaged as single .safetensors files for the ComfyUI-InstantHMR node pack. Each file holds the network and the MHR body rig it was trained with, so it loads with no other model. The output is the same MHR body as ComfyUI's native SAM 3D Body nodes, ready for their smoothing, rendering and GLB / BVH export.

Put the files in ComfyUI/models/detection and pick one in Load InstantHMR Model.

File	Backbone	Parameters	3DPW PA-MPJPE (mm)	Licence
instanthmr_nano.safetensors	HGNetV2-B0	20.0 M	46.6	SAM License, non-commercial
instanthmr_tiny.safetensors	HGNetV2-B1	20.3 M	44.0	SAM License, non-commercial
instanthmr_small.safetensors	HGNetV2-B2	24.4 M	41.7	SAM License, non-commercial
instanthmr_medium.safetensors	HGNetV2-B4	32.2 M	40.0	SAM License, non-commercial
instanthmr_medium_commercial.safetensors	HGNetV2-B4	32.2 M	50.7	SAM License

Licence

All five files are under the SAM License, following the upstream release (momolesang/InstantHMR): they are trained on SAM 3D Body annotations and contain parts of its checkpoint.

• nano, tiny, small, medium: non-commercial use only, because some of their training images are licensed for research or non-commercial use only.
• medium_commercial: trained only on images that allow commercial use, so commercial use is allowed under the SAM License terms.
• The MHR body rig inside every file comes from Meta's MHR (Apache-2.0).

Conversion

Converted from the official ONNX graphs with scripts/convert_onnx.py in the node pack, which checks the result against the ONNX graph with onnxruntime (outputs match to about 1e-5).

Credits

InstantHMR by Mohamed Adjel, Kahina Chalabi and Vincent Bonnet (paper, code). MHR body model by Meta.

---

中文

InstantHMR 的权重，重新打包成单个 .safetensors 文件，供 ComfyUI-InstantHMR 节点包使用。每个文件都包含网络和它训练时用的 MHR 人体骨架，不需要额外的模型。输出的 MHR 人体和 ComfyUI 内置 SAM 3D Body 节点一致，可以直接接它们的平滑、渲染和 GLB / BVH 导出。

把文件放到 ComfyUI/models/detection，在 Load InstantHMR Model 里选择即可。

文件	主干网络	参数量	3DPW PA-MPJPE (mm)	许可证
instanthmr_nano.safetensors	HGNetV2-B0	2000 万	46.6	SAM License，仅限非商用
instanthmr_tiny.safetensors	HGNetV2-B1	2030 万	44.0	SAM License，仅限非商用
instanthmr_small.safetensors	HGNetV2-B2	2440 万	41.7	SAM License，仅限非商用
instanthmr_medium.safetensors	HGNetV2-B4	3220 万	40.0	SAM License，仅限非商用
instanthmr_medium_commercial.safetensors	HGNetV2-B4	3220 万	50.7	SAM License

许可证

和上游发布（momolesang/InstantHMR）一致，五个文件都采用 SAM License：它们用 SAM 3D Body 的标注训练，并包含它的部分模型参数。

• nano、tiny、small、medium：只能非商用，因为部分训练图片只允许研究或非商业用途。
• medium_commercial：只用允许商用的图片训练，按 SAM License 的条款可以商用。
• 每个文件里的 MHR 人体骨架来自 Meta 的 MHR（Apache-2.0）。

转换方式

用节点包里的 scripts/convert_onnx.py 从官方 ONNX 计算图转换，并用 onnxruntime 和 ONNX 计算图逐项比对（输出误差约 1e-5）。

致谢

InstantHMR 作者 Mohamed Adjel、Kahina Chalabi、Vincent Bonnet（论文、代码）。MHR 人体模型来自 Meta。