Motion Generation Tool Tests

A research-evaluation / compatibility-testing repo, not a production tool. Open Source Films (OSF) used a RunPod RTX 5090 pod (Blackwell, CUDA 12.8) to install, patch, and run three published NVIDIA human-motion-generation research projects end to end, and recorded exactly what it took to get each one working — install order, patched dependency versions, gated-asset requirements, and upstream bugs hit and fixed along the way.

Runpod GPU credit made this evaluation possible. RTX 5090 / Blackwell support in the wider PyTorch/CUDA ecosystem is still new enough that "pip install and run" routinely fails against a project's stated requirements — every fix documented in INSTALL_NOTES.md (Blackwell PyTorch reinstalls, a missing-EGL-library fix, a GVHMR upstream-bug workaround) came out of that process.

What's here

• INSTALL_NOTES.md — the real, warts-and-all install log for all three tools: exact commands, every snag hit, and how each was resolved.
• launch_kimodo.sh, launch_stmc.sh, run_genmo_example.sh, run_smoke_tests.sh — the launcher/test harness scripts OSF wrote around each tool (RunPod-proxy port wiring, env setup, smoke-test entry points). These are OSF's own code.
• envs/common_env.sh — points HF/pip/uv/torch caches at the project directory instead of the container filesystem (relevant on RunPod: only the container disk is ephemeral, a mounted persistent volume is not).
• outputs/kimodo/smoke_test_walk_wave.npz — the actual output of a successful Kimodo smoke test (150 frames of SOMA-skeleton motion data: joint positions, rotation matrices, root trajectory, foot contacts — pure numeric arrays, verified to contain nothing beyond the documented motion-data schema). This is generated sample data, not a model checkpoint. OSF licenses any rights it holds in the sample under the repository MIT terms; the NVIDIA model licence governs the checkpoint separately and states NVIDIA does not claim ownership of generated outputs. See the licensing details.

What's not here, and why

This repo deliberately does not include:

• The three upstream tool repos (Kimodo, GENMO/GEM, STMC) — clone them yourself from the links below; each keeps its own upstream license (see THIRD_PARTY_LICENSES.md).
• Any model weights or checkpoints — Kimodo-SOMA-RP-v1, the GEM checkpoint, HMR2/ViTPose, STMC's pretrained models. Fetch these from the sources documented in INSTALL_NOTES.md.
• SMPL / SMPL-X / MANO / SMPL+H body-model assets — these require individual registration at the Max Planck Institute's clickthrough-license portals (smpl-x.is.tue.mpg.de, mano.is.tue.mpg.de) and are not ours to redistribute.
• GENMO or STMC generated output files — GENMO's output is derived from an identifiable person's test footage (see below); STMC's output is derived from an explicitly non-commercial pretrained model. Both are described in INSTALL_NOTES.md in words instead: what was generated, what pipeline stages ran, and what the result looked like.
• The test source video used for the GENMO run — a short clip trimmed from OSF's own behind-the-scenes production footage that visibly shows a real person's face and likeness. Neither that clip nor anything rendered from it is included.
• Environments, caches, logs, or any machine-specific state from the pod this was run on.

The result: what's published here is OSF's own harness/documentation, one genuinely clean output artifact, and a factual account of what worked — not a redistribution of anyone else's restricted model, weights, or research assets.

The three tools tested

Tool	Upstream	What it does	Code license	Model/weights license
Kimodo	nv-tlabs/kimodo	Text/constraint-conditioned motion diffusion	Apache-2.0	Kimodo-SOMA-RP-v1: NVIDIA Open Model License (commercial use permitted)
GENMO / GEM	NVlabs/GENMO	Generalist motion model (video/text/audio → SMPL)	NVIDIA OneWay Noncommercial License	Not independently re-verified beyond the code license — see THIRD_PARTY_LICENSES.md
STMC	nv-tlabs/stmc	Timeline-based (Spatio-Temporal Motion Collage) motion generation	Explicitly non-commercial (code, MTT dataset, and pretrained models all carry their own upstream restrictions)	See upstream repo

Full detail, sources, and how these verdicts were reached: THIRD_PARTY_LICENSES.md.

Results

All three tools were installed from scratch and produced a successful end-to-end generation on this pod:

• Kimodo — text prompt → 150-frame walk-and-wave motion on a 77-joint SOMA skeleton. Output included in this repo (outputs/kimodo/smoke_test_walk_wave.npz).
• GENMO / GEM — video-conditioned motion capture: a short test clip → YOLOv8 person detection → ViTPose keypoints → HMR2 features → GEM diffusion → SMPL parameters, rendered as in-camera, global, and side-by-side comparison videos. Verified by visual review (the SMPL mesh overlay tracked the real person's position and pose correctly) — result described in INSTALL_NOTES.md, not redistributed (see above).
• STMC — timed multi-track text timeline ("walking in a circle" + overlapping "raising both arms in the air") → joint positions, SMPL parameters, and rendered skeleton/mesh videos via a non-commercial pretrained checkpoint. Verified by visual review — result described in INSTALL_NOTES.md, not redistributed.

See INSTALL_NOTES.md for exact commands, every gated asset each tool required, and the real upstream bugs hit and fixed along the way (a missing-body-model-file GENMO issue traced to an un-vendored GVHMR dependency, a headless-EGL rendering fix for STMC, a Blackwell/cu128 PyTorch reinstall forced by an unpinned transitive dependency).

Running this yourself

1. Clone the three upstream repos into repos/ (see links above).
2. Follow INSTALL_NOTES.md for each tool's environment setup and gated-asset requirements.
3. source envs/common_env.sh before any install/run step.
4. Use launch_kimodo.sh / launch_stmc.sh / run_genmo_example.sh / run_smoke_tests.sh as documented in the "What's here" section above.

See RUNPOD.md for pod-specific GPU/proxy notes.

The CLI wrappers run_smoke_tests.sh and run_genmo_example.sh now create their own logs/ directory and propagate a failing model-command or log-write exit status rather than printing a misleading completion message. In particular, run_smoke_tests.sh all fails if its Kimodo test fails; STMC does not claim outputs after a failed generation. The all mode still runs only Kimodo, not GENMO/STMC.

STMC Gradio sharing guard: launch_stmc.sh runs envs/disable_stmc_share.py before starting app.py. It accepts only one simple, module-level SHARE = True or SHARE = False declaration in the upstream repos/stmc/app_config.py. It changes True to False atomically, leaves an existing False unchanged, and refuses to launch if the declaration is absent, ambiguous, dynamic or otherwise changed by upstream. The former unverified sed replacement could silently miss spacing changes and start a demo with a public Gradio share link enabled. This safeguard only checks the intended upstream configuration variable: it cannot prove that a later STMC version doesn't create other public links. The RunPod HTTP proxy remains independently reachable when configured; no authentication is added by this guard. Use a trusted access-controlled deployment environment rather than treating SHARE=False as a firewall.

Offline publication checks

The repository's public GitHub Actions workflow runs six CPU-only sample/schema and shell-syntax tests, six smoke-launcher behaviour tests and eight STMC sharing-guard tests. The first group verifies the already published outputs/kimodo/smoke_test_walk_wave.npz artifact: seven numeric/bool fields, expected dtypes and dimensions, consistent 150-frame length, finite floating-point values, no object/pickle arrays, and 77-joint motion geometry. The launcher and sharing-guard tests run copied shell wrappers only against isolated fake binaries and disposable upstream configuration files to verify success/failure reporting, log creation, sharing deactivation and fail-closed behaviour without invoking any motion-generation model.

Only NumPy is installed for the artifact checks; launcher tests use Python's standard library. No GPU, Kimodo/GENMO/STMC model, weights, human subject imagery or RunPod service is required. A passing result does not validate model installation, launchers in a current GPU environment, kinematic realism, motion quality or any upstream licences. The captured output artifact and original research verdicts are unchanged.

Status

Research evaluation / compatibility testing, run once to completion on 2026-07 against a specific pod configuration. Not maintained as a production tool — treat INSTALL_NOTES.md as a point-in-time account of what worked, not a guarantee against future upstream changes.

---

Runpod-sponsored GPU compute. Open Source Films.