xiaoerofficial021-collab/FlowHMR-Windows

What it is

FlowHMR-Windows is a desktop distribution of the FlowHMR research project, packaged for Windows. It turns ordinary video into 3D human motion, offers a 3D preview, and exports the result as FBX. The upstream model is a 0.46B-parameter flow-matching system trained for physically plausible motion recovery; the released flowhmr_latest checkpoint adds GRPO post-training with a physics reward and is the recommended model (model zoo).

Beyond single-person capture, the desktop app supports selecting multiple people, synchronized preview, original-audio playback, finger refinement, contact correction, and multi-person FBX export. A companion Blender add-on, FlowHMR 2 MMR, retargets the exported FBX onto MMR rigs while preserving each performer's entry timing and relative position.

It is aimed at animators, mocap hobbyists, and researchers who want local video-to-motion capture without setting up the full Python pipeline, and at Blender users who need motion for MMR rigs.

Quick start

1. Download FlowHMR-Windows.zip from the latest release and unzip it; the base package includes Python, the UI, and FBX export.
2. Launch FlowHMR.exe.
3. Open 运行环境 (Environment) in the top-right corner and download/enable a CUDA version.
4. Under 模型 (Models), download a public model.
5. Register on the MANO download page, get the Extended SMPL+H model (smplh.tar.xz), and extract it directly into models/body_models/smplh/. For finger refinement or body-shape switching, also download Models & Code (mano_v1_2.zip) and extract it into models/.
6. Import a video, select the people to reconstruct, preview the motion, and export FBX.

Optional: install the FlowHMR 2 MMR add-on ZIP via Blender's Edit → Preferences → Add-ons (Blender 3.6+), press N in the 3D viewport, and open the FlowHMR 2 MMR tab. It requires an MMR control rig generated beforehand; see the release notes.

Constraints

• The app needs a CUDA-capable GPU environment downloaded through the built-in environment manager; models are loaded only when a generation runs.
• SMPL-H and MANO body models require accepting the MPI license and cannot be bundled; the app will not work without them.
• Source building and detailed Windows usage are covered in the Windows guide, with model directory layout in models/README.
• The full research pipeline (training, GRPO post-training, web demo) is separate: V2M setup uses Python 3.10 + PyTorch 2.5.1 (cu118); the physics pipeline needs Isaac Gym Preview 4 with Python 3.8 + PyTorch 1.13 (cu117), communicating through /dev/shm. Training data is not bundled, and Wild-4K releases only clip IDs, not videos.
• Code is Apache-2.0; model weights follow the license on the Hugging Face model card.

HotRepo's take

The repository entered HotRepo's Python new-project list on 2026-11-11 with 108 stars and 7 net new stars over a roughly 2.5-hour window, and 12 forks. There is no GitHub daily-rank record or Hacker News discussion to draw on. The growth is recent and short-window, so it does not establish sustained adoption. The notable technical fact is the split between a ready-to-run Windows package and a research stack with its own environment requirements; the packaging is what makes the project approachable, but it also means the desktop binary and the upstream paper code are distributed differently and should be treated as separate artifacts.