Blendcap/blendcap-onnx-exports

BlendCap ONNX Exports

Pre-generated ONNX Runtime files for the BlendCap performance capture addon for Blender.

These are NOT standalone models. They're component exports (backbone, FOV encoder, and pose detector) used inside BlendCap's larger SAM 3D Body + YOLO11 + MoGe inference pipeline. You almost certainly want BlendCap itself, not these files directly.

Files

File	Size	Source model	License
backbone_dinov3_fp16.onnx (+ .onnx_data)	668 KB + 1.68 GB	DINOv3 ViT-H backbone from SAM 3D Body	SAM License (see LICENSE)
moge_s_encoder_fp16.onnx	46 MB	MoGe-2 ViT-S encoder	MIT (see LICENSE_MIT.txt)
yolo11m_pose.onnx	40 MB	YOLO11m-pose person detector (Ultralytics)	AGPL-3.0 (see LICENSE_AGPL.txt)

The backbone uses ONNX's external-data format because its weights exceed protobuf's 2 GB in-memory limit. Both backbone_dinov3_fp16.onnx and backbone_dinov3_fp16.onnx_data must be downloaded together and kept in the same directory with their exact names preserved — the .onnx file references .onnx_data by relative filename. Loading fails if either is missing.

yolo11m_pose.onnx is an ONNX export of Ultralytics YOLO11m-pose and is licensed under AGPL-3.0. Its corresponding source is available via the open-source BlendCap repository at https://github.com/Arcomade/BlendCap, satisfying the AGPL source-availability requirement. No commercial Ultralytics license is required.

Usage

Loaded automatically by BlendCap when USE_ORT=1 is set. If you're using these directly:

import onnxruntime as ort

session = ort.InferenceSession(
    "backbone_dinov3_fp16.onnx",
    providers=["CUDAExecutionProvider", "DmlExecutionProvider",
               "CoreMLExecutionProvider", "CPUExecutionProvider"],
)

ORT picks the first available execution provider from the list.

Source

• DINOv3 backbone: derived from facebook/sam-3d-body-dinov3 (gated — see Meta's HF for original weights and license acceptance).
• MoGe-2 encoder: derived from Ruicheng/moge-2-vits-normal.
• YOLO11m-pose detector: derived from Ultralytics YOLO11 (AGPL-3.0).
• Exported via the scripts in BlendCap's source repo.