Native Windows video motion capture with ExMR, FlowHMR and GVHMR, contact correction, humanoid retargeting, and FBX/VMD export.
AI model guides
Choose a model in AI Motion Model before preparing its files or opening videos. ExMR is the default for new work; reopening a project restores its saved selection. Only implemented tools appear in the list.
| Model | Input and body | Fingers | Detailed guide |
|---|---|---|---|
| ExMR | One video; SMPL-X | Included | ExMR |
| FlowHMR | One video; SMPL-H | Included | FlowHMR |
| GVHMR | One video; SMPL-X | Optional HandFlow | GVHMR |
| MAMMA | Two or more synchronized fixed-camera videos; SMPL-X | Included | MAMMA |
| PromptHMR | One video; fixed or optional moving camera; jointly reconstructed people; SMPL-X | Not generated | PromptHMR |
| PEAR | Experimental real-time tracking from webcam or video input; one person, one or multiple fixed views; may be unstable | Included | PEAR |
One-video input does not mean one-person output. ExMR, FlowHMR and GVHMR can
process checked people independently after Find person. Without discovery,
automatic tracking selects one person and may switch identity during overlap.
MAMMA detects and matches people across views during reconstruction; its
Find person button is hidden.
PromptHMR also tracks people during reconstruction. It defaults to two people
and Interaction Prompt on; see its guide for setup and fixed-camera limits.
For manual MAMMA cameras, see the calibration format and JSON example.
Automatic MAMMA cameras combine VGGT-Omega placement with GeoCalib background
vertical direction before 3D fitting; both checkpoints appear in its Model Manager.
GeoCalib reuses the base Flow support overlay rather than MAMMA's separate runtime.
PEAR support is prepared by install.cmd and update.cmd over the shared runtime;
its weights/body data are obtained separately. It is a live preview without motion
recording, offline contact correction, retargeting or export. Camera Align is
attempted first; if it fails, tracking continues from the selected camera with a
warning, without camera fusion or aligned live foot lock.
In this project's subjective tests, ExMR was generally fast with good quality; FlowHMR was slower with the best overall quality; GVHMR was less consistent overall but sometimes better for unusual poses such as backflips. These are not benchmark rankings. PromptHMR could improve relative placement and interactions for multiple people in one video, but its overall motion quality was lower than ExMR in our subjective tests, and it has no finger tracking. MAMMA is a separate multi-view workflow rather than another single-video option.
Installation and shared workflow
- Run
install.cmdin the app folder. It prepares ExMR, FlowHMR and GVHMR runtimes and their supported add-ons, PEAR's shared support, Blender, FFmpeg and native motion correction. It does not download AI weights or registered body data. - For MAMMA, additionally follow its runtime setup.
Setup_MAMMA.cmdis separate from the base installation and update commands. PromptHMR independently uses Setup_PromptHMR.cmd and does not require MAMMA/SDF. Both optional setups require the documented C++/CUDA build tools and provider license review. Close the app before setup; reopen it before preparing weights.update.cmddoes not install either optional runtime; rerun its own setup if the app requests it after an update. - Run
RGBMoCap.cmd, select the model and use Model Manager. Follow each row's exact download/import instructions. Some files require registration and license acceptance; Google Drive downloads may need manual import. - Open footage, trim, reconstruct and review. Contact correction defaults on; Foot defaults on and Hand off. Edit contact timings and Apply Correction without rerunning AI inference. File > Export... offers FBX, BVH or a portable Blender project for the displayed applied motion. MAMMA also offers checked bodies and all cameras in one Blender scene; see export options.
For PEAR, use Start Live Tracking instead of reconstruction. It displays live bones only; offline contact, project-save, retarget and export steps do not apply. Multiple synchronized files/webcam slots and Camera Align are described in its guide, along with filtering, World Position and Reset Bones.
Model-specific settings affect reconstruction; they are separate from contact correction and display-only viewport controls. Original and corrected results are preserved. One project can hold separate results for each model. MAMMA saves all of its camera videos with Save Project As.
Use Resource Mode > Automatic initially. Resource Safety, default on, controls low-RAM/Windows headroom stopping and CUDA allocator limits. Off bypasses these limits; adaptive batching and actual allocation-error handling stay active. See Resource Safety. More frames, pixels, people or views can increase work; no setting guarantees that every clip fits. Automatic selects supported Low Memory paths when the GPU budget is low; a supported operation can recover once after CUDA OOM. Low Memory uses CPU loading/offloading and existing bounded image batches, which can be slower and need more RAM. It does not reduce model precision, motion frame count or MAMMA view count/iterations. Decisions appear in Console and the CMD window. PromptHMR keeps its validated one-image batch in both modes; additional block offloading is not implemented. Low Memory does not guarantee a smaller PromptHMR footprint. See resources and measured validation scope.
Low Memory primarily saves VRAM, not system RAM. Short paired mode checks measured ExMR at 5.68 → 3.87 GiB and FlowHMR at 6.10 → 2.77 GiB allocated VRAM; GVHMR's short test had no meaningful whole-job reduction. MAMMA's paired whole-pipeline VRAM saving was not recorded; only its shared VGGT offload component has a controlled comparison. See measured savings and speedups for RAM, timing, test scope and numerical differences. These figures are not minimum requirements or predictions for every clip.
See the user guide for shared controls, contacts, retargeting, FBX/VMD export, storage and cleanup. For agent-assisted setup and diagnostics, use the agent guide. Model code, weights and body data retain their own licenses.