Skip to content

GenMM

Basic

Development line: project:genmm · thread genmm-development
Last event: 2023-06-02 · 1 dated since 2023-06-02 · Researched: 2026-09-04 · confidence: medium

What it is

GenMM — a SIGGRAPH 2023 research codebase for technical animators who need new skeletal-motion sequences from one or a few example clips, not a text-to-motion model.

  • Synthesizes diverse variants, completes motion, follows keyframe poses, creates loops, reassembles motion across heterogeneous skeletons, and changes locomotion trajectories.
  • The documented local stack is Python 3.8, PyTorch 1.12.1 and CUDA 11.3; the authors report about 0.05 s per motion on a Tesla V100 and 0.2 s on an Apple M1.

Development line

  • 2023-06-02 — GenMM public project materials linked. On 2023-06-02, the project published its project page, source repository, and demo link. This milestone opened evaluation entry points for the project, though the unavailable source text leaves the exact release claim and technical details unestablished.

What changed

2023-06-02 — GenMM went live with a project page, source repository, and browser demo as an example-based, training-free motion-synthesis system. The page documents motion completion, keyframe guidance, looping, reassembly, and locomotion. 2023-07-13 — the entry provides only the existing browser-demo endpoint. It identifies no new model, package version, or feature, so this is not a confirmed software release. 2026-09-04 — the original GitHub URL redirects to liyuuuu98/GenMM; the current releases page exposes a single Blender add-on pre-release, v0.0.1. The retrieved release page gives only "29 Jul 12:48", not a year, so no full release date is assigned.

How to use this

From 2023-06-02, we could use the public GenMM project page, source repository, and demo as entry points for evaluation; the 2023-07-13 demo link alone does not establish a further practical change.

  1. Clone the repository and reproduce its documented environment: Python 3.8, PyTorch 1.12.1, CUDA 11.3, docker/requirements.txt and torch-scatter==2.1.1. — https://github.com/wyysf-98/GenMM
  2. Provide a BVH motion clip and run python run_random_generation.py -i ; use run_random_generation.py for further synthesis configuration. — https://github.com/wyysf-98/GenMM
  3. For Blender, install the listed Python packages into Blender's bundled Python, install the add-on ZIP through Preferences > Add-ons > Install, then select an armature and open the GenMM panel. — https://github.com/wyysf-98/GenMM
  4. For the web demo, select an example with Next, use Generate and Controls, then set frames, noise_sigma, loop, patch_size, coarse_ratio, pyr_factor and optimisation steps. — https://github.com/wyysf-98/GenMM_demo
  5. If the shared demo waits indefinitely, duplicate its Hugging Face Space, obtain your own API URL and replace api_url in Controls. — https://github.com/wyysf-98/GenMM_demo

Best practices

  • Pin or containerize the documented Python, PyTorch and CUDA versions before changing dependencies. The official instructions do not provide a newer compatibility matrix. — https://github.com/wyysf-98/GenMM
  • Validate final motion in a rig-aware workflow, not only in the browser viewer. The project page warns that its Three.js FBX parser has no foot-contact fix and may create artifacts. — https://weiyuli.xyz/GenMM/
  • When the shared demo is slow, use a duplicated Space with your own API URL rather than repeatedly retrying the CPU-backed shared endpoint. — https://github.com/wyysf-98/GenMM_demo

Superseded by this

  • 2023-06-02 — the README release checklist is obsolete as an availability statement: the repository now contains the local runner and GitHub lists the Blender add-on v0.0.1 pre-release.
  • 2023-07-13 — a browser-demo-only workflow is obsolete for repeated work: first-party documentation now provides a copied-Space/private-api recovery path and warns of viewer artifacts and shared-CPU stalls.

Still unknown

  • No first-party source read today provides a 2026-tested Python, PyTorch, CUDA or Blender compatibility matrix.
  • The 2023-07-13 entry supplies only the existing demo URL. Without its source text, we cannot attribute any separate capability or version to that date.
  • The current release page displays v0.0.1 as a pre-release and gives 29 Jul 12:48, but the retrieved view does not expose the year.
  • The browser demo requires JavaScript; live end-to-end generation was not verified in this research.

Sources

source title read
https://weiyuli.xyz/GenMM/ Example-based Motion Synthesis via Generative Motion Matching 2026-09-04
https://github.com/wyysf-98/GenMM GitHub — liyuuuu98/GenMM 2026-09-04
https://weiyuli.xyz/GenMM_demo/ patch-based_motion_synthesis 2026-09-04
https://github.com/wyysf-98/GenMM_demo GitHub — liyuuuu98/GenMM_demo 2026-09-04
https://github.com/wyysf-98/GenMM/releases GitHub — GenMM releases 2026-09-04

Agent brief

  • Subject: project:genmm, thread genmm-development, 1 dated events 2023-06-02 → 2023-06-02.
  • Practical note: From 2023-06-02, practitioners could use the public GenMM project page, source repository, and demo as the entry points for evaluation; the 2023-07-13 demo link alone does not establish a further practical change.
  • Confidence: medium. Dated supersedes above are the authority for what is obsolete.