CoolFace
Apppublic

nvidia/synthda-demo

sourceHugging Facefair-noncommercial-research-licenseupdated 1d agoView on Hugging Face
3likes
App README

SynthDA Motion Playground

An interactive viewer for actual saved SynthDA poses. Compare two input motions with an augmented output, inspect a lineup of five variations, and follow a source RGB clip through pose extraction and a fitted SOMA human visualization.

Open the demo · Research code and publications

[image]

Included in this release

  • —Video → extracted pose → human figure: one source video drives both the extracted skeleton and a fitted native SOMA body. Toggle the body and joint overlay.
  • —Motion playground: Compare motions, Variation lineup and Overlay poses; saved-variant selection, full camera rotation, zoom, frame scrubbing, playback speed, SOMA/Skeleton display, Original/Smoothed preview and original-pose download.
  • —Variation gallery: three verified fall pose examples, two short motion studies, and all 33 videos from the original Hugging Face Space.
  • —SOMA bodies: 33 fitted motions (six sources and 27 variants), rendered locally in WebGL 2 with the native SOMA-X neutral body. Hands stay neutral; fitted surfaces approximate the saved joints and do not reconstruct the original person.
  • —Additional examples and sourced recognition-result records can be added through the manifest.

This release contains one action class: falling. Multiple fall examples are not represented as multiple action classes. There are no fabricated recognition scores or motion-quality measurements. This is saved-result visualization, not a live inference endpoint.

Run locally

No Node install, build step, GPU or backend is required for the viewer.

sh
python3 -m http.server 7860 --bind 127.0.0.1

Open http://localhost:7860. Serve the folder over HTTP; opening index.html directly from the filesystem will block JSON loading in many browsers.

Publish or extend

See DEPLOYMENT.md for Hugging Face and GitHub instructions, ASSET_FORMAT.md for additional examples and meshes, and ASSET_PROVENANCE.md for source revisions and limitations.

The Space's pre-existing license metadata is preserved above. Upstream repository material retains its NSCLv2 notice; this viewer does not relicense upstream assets. Publicly hosting saved visualizations does not change the license terms of the generation pipeline or source assets.

The SOMA-derived assets retain their Apache-2.0 license and notice. Three.js retains its MIT license. See SOMA.md for fitting, rendering and reproduction details.