Instructions to use fernandotonon/QtMeshEditor-sam3dbody-onnx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- SAM 3D Body
How to use fernandotonon/QtMeshEditor-sam3dbody-onnx with SAM 3D Body:
import cv2 import numpy as np from notebook.utils import setup_sam_3d_body from tools.vis_utils import visualize_sample_together estimator = setup_sam_3d_body("fernandotonon/QtMeshEditor-sam3dbody-onnx") img_bgr = cv2.imread("path/to/image.jpg") outputs = estimator.process_one_image(cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)) rend_img = visualize_sample_together(img_bgr, outputs, estimator.faces) cv2.imwrite("output.jpg", rend_img.astype(np.uint8)) - Notebooks
- Google Colab
- Kaggle
SAM 3D Body (body branch) + MHR β ONNX, for QtMeshEditor
ONNX export of Meta's SAM 3D Body (facebook/sam-3d-body-dinov3, body-only inference path) and of the
MHR (Momentum Human Rig) body model, as used by QtMeshEditor
for qtmesh body photo.jpg -o body.obj β a human body mesh (18,439 vertices) from ONE photo.
This repository is the standalone model card + mirror. The application downloads from the aggregate repo
fernandotonon/QtMeshEditor-models under sam3dbody/; both hold identical files.
Files
| file | what | licence |
|---|---|---|
sam3d_backbone.onnx + .onnx.data |
DINOv3-ViT-H+/16 backbone, fp32, image [1,3,512,512] (RGB 0..1) β features [1,1280,32,32] |
SAM License |
sam3d_decoder.onnx |
camera-ray conditioning + promptable decoder (6 layers, intermediate pose feedback) + pose/camera heads | SAM License |
mhr_body.onnx |
MHR body forward (shape [1,45] + model params [1,204] β vertices in cm, joint translations/rotations), no pose correctives, no face model |
Apache-2.0 (MHR data) |
mhr_faces.i32 |
MHR body topology, 36,874 triangles, int32 little-endian | Apache-2.0 (MHR data) |
mhr_uv.bin |
MHR UV layout (int32 nt, nf; float32 texcoords[nt,2]; int32 texcoord_faces[nf,3]) for texturing | Apache-2.0 (MHR data) |
LICENSE-SAM.txt, LICENSE-MHR.txt |
the licence texts β the SAM License requires a copy to travel with the SAM materials |
Provenance and fidelity
Exported by scripts/export-sam3dbody-onnx.py in the QtMeshEditor repository from Meta's checkpoint (the script
imports Meta's own package from a local clone; no Meta source files are redistributed). Every graph was checked
against Meta's PyTorch run on the same inputs: decoder outputs agree to ~1e-6, backbone to ~2e-5 (fp32 here vs bf16
in Meta's code), MHR to ~1e-4 cm. What is deliberately not reproduced: MHR pose correctives (a 664 MB network;
a posed body moves by a mean 2 mm / max 25 mm without it) and the face-expression model (the body head zeroes it).
Pre-processing (person crop, affine warp, CLIFF condition, camera rays) lives in the application
(src/BodyFromPhoto/SamBodyPredictor.cpp), not in the graphs.
Licence β read before using
sam3d_backbone.onnx*andsam3d_decoder.onnxare derivative works of Meta's SAM Materials and are distributed under the SAM License (LICENSE-SAM.txt; Meta Platforms, Last Updated November 19, 2025). It is not an OSI open-source licence: it carries an indemnity clause, a litigation-termination clause, trade-control restrictions and Meta's right to update the terms. By using these files you accept it.mhr_body.onnxandmhr_faces.i32derive from Meta's MHR release data, which is Apache-2.0 (LICENSE-MHR.txt).- Hosting here follows the maintainer's own reading of the SAM License and is not legal advice.
Not affiliated with or endorsed by Meta.
Model tree for fernandotonon/QtMeshEditor-sam3dbody-onnx
Base model
facebook/sam-3d-body-dinov3