m-hamza-mughal/miburi

MIBURI — live demo

Implementation demo for MIBURI: Towards Expressive Interactive Gesture Synthesis (CVPR 2026) — paper · project page · code.

Talk into your microphone; a frozen Moshi 7B answers in speech while MIBURI's Gesture LM drives a SMPL-X character's full body, hands and face at 25 fps in the viewer next to it.

Headphones are recommended — without them your microphone picks up the model's own voice and it starts replying to itself.

Demo availability

The demo needs a dedicated ~24 GB GPU and is powered on for scheduled demo windows rather than left running. When it is off, this URL serves a landing page with the paper, video results and code. When it is on, the same URL is the full dashboard.

To run it yourself on any 24 GB GPU, see the repo README — miburi-demo is the same code this Space runs.

How this Space is built

Dockerfile installs the miburi package from GitHub and bakes every weight into the image at build time (build time is unbilled; GPU runtime is not). entrypoint.sh then picks one of two modes based on $ACCELERATOR:

The demo normally wants two ports (the aiohttp dashboard and Viser's own server). Spaces expose one, so --embed-viser serves Viser's single-file client bundle at /viser and proxies its WebSocket through the same port.

Licensing

MIBURI code and weights are CC BY-NC 4.0 — non-commercial use only. Third-party components keep their own licenses: Moshi/Mimi (MIT code, CC-BY-4.0 weights, Kyutai), SMPL-X (Max Planck Society, non-commercial), Mixamo characters (Adobe Mixamo terms).

@InProceedings{mughal2026miburi,
    title = {MIBURI: Towards Expressive Interactive Gesture Synthesis},
    author = {M. Hamza Mughal and Rishabh Dabral and Vera Demberg and Christian Theobalt},
    booktitle = {Computer Vision and Pattern Recognition (CVPR)},
    year = {2026}
}