InstantHMR

InstantHMR, by Naturalpad and LAAS-CNRS, is a real-time human mesh recovery model. For each person box it returns MHR body pose and identity parameters, 70 keypoints in 2D and 3D, the camera translation and an 18,439-vertex mesh. LibreYOLO ports it natively in PyTorch and finds the people with RF-DETR.

Tasks
body mesh
Install
pip install libreyolo
Support tier
Supported, since v1.8.0. Supporting trainables: kept green in CI, features land opportunistically.
Upstream
InstantHMR by Naturalpad and LAAS-CNRS, CC-BY-NC-4.0 (n, t, s, m checkpoints); Apache-2.0 (m commercial checkpoint). Paper, source
Licenses
Code Apache-2.0, weights CC-BY-NC-4.0 (n, t, s, m); Apache-2.0 (m-commercial). Commercial use

Install

InstantHMR needs no optional extra. The default person detector, LibreRFDETRm.pt, needs the RF-DETR extra.

bash
pip install "libreyolo[rfdetr]"

Predict

Weights download from the LibreYOLO Hugging Face mirrors on first use, after a licence notice.

Python
from libreyolo import LibreYOLO model = LibreYOLO("LibreInstantHMRm-mesh.pt")   # people found with LibreRFDETRm.ptresult = model("people.jpg") meshes = result.meshes            # one row per box in result.boxesprint(meshes.body_model)          # "mhr"print(meshes.vertices.shape)      # (people, 18439, 3) metres, camera frameprint(meshes.joints2d.shape)      # (people, 70, 2) pixels
Your own boxes
from libreyolo import LibreYOLO model = LibreYOLO("LibreInstantHMRm-mesh-commercial.pt")boxes = [[40, 20, 260, 600]]      # x1, y1, x2, y2 per personresult = model("people.jpg", person_boxes=boxes)

People are found by an RF-DETR detector (LibreRFDETRm.pt by default). Pass person_boxes= to use your own boxes, or person_detector= to swap the detector. result.meshes rows line up with result.boxes. Everything is in the camera frame of the original image: transl is in metres, vertices and joints3d are metric and include transl, and joints2d is in pixels on the original image. global_orient, body_pose and betas hold the MHR parameters; MHR rotations are Euler angles in radians. See the body mesh task page for the full payload.

Variants

Four sizes, n, t, s and m, all at a 288 px person crop, plus LibreInstantHMRm-mesh-commercial.pt, an m trained only on commercially usable images. Upstream publishes ONNX only; LibreYOLO rebuilds PyTorch checkpoints that match the upstream ONNX graphs within 8e-5 for all five releases.

Train

train() fine-tunes a pretrained checkpoint on a YOLO-pose dataset with COCO-17 person keypoints, using 2D keypoint, reprojection and frozen-prior losses. Training from scratch and resume are not supported. val() raises: the usual body-mesh benchmarks (3DPW, EMDB) are research-license only, so none is bundled; the MPJPE, PA-MPJPE and PVE metrics are in libreyolo.validation.mesh_metrics for a dataset you hold.

Python
from libreyolo import LibreYOLO # Fine-tune a pretrained checkpoint on a YOLO-pose COCO-17 dataset.model = LibreYOLO("LibreInstantHMRs-mesh.pt")model.train(data="my-pose-dataset.yaml", epochs=20)

Export

TaskONNXTorchScriptExecuTorchTensorRTOpenVINOPaddleMNNRKNNAxelerancnnTFLiteCoreMLCore AI
Body meshBody mesh to ONNX: supportedBody mesh to TorchScript: not supportedBody mesh to ExecuTorch: not supportedBody mesh to TensorRT: not supportedBody mesh to OpenVINO: not supportedBody mesh to Paddle: not supportedBody mesh to MNN: not supportedBody mesh to RKNN: not supportedBody mesh to Axelera: not supportedBody mesh to ncnn: not supportedBody mesh to TFLite: not supportedBody mesh to CoreML: not supportedBody mesh to Core AI: not supported

export(format="onnx") writes a graph with image and cliff_cond inputs and the mesh faces in its metadata.

Python
from libreyolo import LibreYOLO model = LibreYOLO("LibreInstantHMRm-mesh.pt")model.export(format="onnx")   # inputs image, cliff_cond; faces in metadata

Checkpoints

Every file above exists in the LibreYOLO org today and downloads on first use.

Licensing

Check the license on the Hugging Face repository of the specific weights you download. Every checkpoint in the LibreYOLO org carries one, and they are not always the same across a family. That repository is the authoritative source; the summary below describes what applied when this page was last verified.

This is a description of the licenses involved, not legal advice. If the answer matters commercially, read the licenses yourself and take your own counsel.

Original work
InstantHMR, Naturalpad and LAAS-CNRS
Upstream license
CC-BY-NC-4.0 (n, t, s, m checkpoints); Apache-2.0 (m commercial checkpoint)
LibreYOLO code
MIT
Weights
CC-BY-NC-4.0 (n, t, s, m); Apache-2.0 (m-commercial), republished at huggingface.co/LibreYOLO
Interpretation
The licence depends on the checkpoint. The n, t, s and m weights are CC BY-NC 4.0: attribution is required and commercial use is not allowed, because part of their training images (SA-1B, AI Challenger) are non-commercial. LibreInstantHMRm-mesh-commercial was trained on Harmony4D and commercially licensed COCO images only and is Apache-2.0, which allows commercial and closed-source use if you keep the license text and notices. Every checkpoint was also trained on the SAM 3D Body dataset annotations, released by Meta under the SAM License, whose terms pass through with the weights, and they embed the MHR body constants (Apache-2.0). The upstream code is Apache-2.0 (Naturalpad-studio/InstantHMR at commit 5437400d4b25c1972821e51916fac34bd96acce6; LICENSE and NOTICE in libreyolo/models/instanthmr); the mesh decode was written clean-room so that no SAM-licensed sam-3d-body code is used, and the fine-tuning trainer is original LibreYOLO MIT code. Upstream publishes ONNX only (momolesang/InstantHMR, revision 07988f51c8c8908cfc7bb44fc18f645c9551b0bf); LibreYOLO converts them to PyTorch and mirrors them as LibreYOLO/LibreInstantHMR{n,t,s,m}-mesh and -m-mesh-commercial, not gated, and prints a notice before auto-download.

InstantHMR is by Naturalpad and LAAS-CNRS (Naturalpad-studio/InstantHMR). The code is Apache-2.0. The weights licence depends on the checkpoint:

  • n, t, s and m: CC BY-NC 4.0, non-commercial only, because part of their training images (SA-1B, AI Challenger) are non-commercial.
  • m-commercial: Apache-2.0, trained on Harmony4D and commercially licensed COCO images only.

Every checkpoint was also trained on the SAM 3D Body dataset annotations, released by Meta under the SAM License, whose terms pass through with the weights. The mesh decode was written clean-room, so no SAM-licensed sam-3d-body code is used. The checkpoints are mirrored as LibreYOLO/LibreInstantHMR{n,t,s,m}-mesh and LibreYOLO/LibreInstantHMRm-mesh-commercial, not gated.

Citation

@misc{adjel2026instanthmr,
  title        = {InstantHMR: Real-Time Human Mesh Recovery on Your Phone},
  author       = {Adjel, Mohamed and Chalabi, Kahina and Bonnet, Vincent},
  year         = {2026},
  howpublished = {HAL preprint hal-05854387},
  url          = {https://hal.science/hal-05854387},
  note         = {LAAS report no. 26486}
}

Copied from the authors' citation block at github.com/Naturalpad-studio/InstantHMR#citation.

Verified against LibreYOLO v1.8.0. Support tables, checkpoints and benchmark numbers on this page are generated from the released library and the published weights, not written by hand.