README.md

metadata

title: Controllable Human Motion Generation
emoji: 🌍
colorFrom: red
colorTo: indigo
sdk: gradio
sdk_version: 6.15.0
python_version: '3.12'
app_file: app.py
pinned: false
license: mit
short_description: Style-SALAD human motion generation on ZeroGPU

Style-SALAD Motion Studio

The interface follows Joao’s files: permanent navy dark mode, lilac accents, Inter body text and Space Grotesk titles. Font files are bundled locally with their SIL Open Font Licenses. Reference upload, generation settings and experimental content motion are grouped under Advanced. Hover or focus the i icons for input tips; reference IDs appear only in Specific ID mode.

Generate stylized human motion with the published Style-SALAD model on Hugging Face ZeroGPU. This deployment runs actual inference in the Space. It does not forward inference to the owner's Windows computer.

Choose one of the 47 released movement styles and sample a reference with a seed, choose a specific ID from that style, or upload raw HumanML3D features (T,263) at 20 fps. References need at least 41 frames and use a centered window of at most 196. Describe an action in English. Each result provides MP4, BVH, NPY, NPZ and JSON.

Features are denormalized with the published Comp_v6_KLD005 statistics and reconstructed into 22 joints. MP4 is drawn from those joints over a colored checkerboard fixed in world coordinates, with a tracking camera and trail. BVH is a separate inverse-kinematics export from the same joints; its fitting error is recorded in the result JSON. MP4 is not rendered from BVH.

The model is loaded at startup using ZeroGPU's CUDA emulation. Only real sampling and reconstruction reserve GPU time. Reference download, validation, MP4 encoding and BVH export run on CPU outside the GPU reservation. Frozen model weights do not disable latent gradients used by style guidance.

The released 47-style catalogue is available through a small verified index. Only the selected feature file is fetched from the official SMooDi archive using HTTP byte ranges, with ZIP CRC and NPY validation. This repository does not mirror datasets or pretrained weights. Generation weights and CLIP are downloaded directly from their official public sources and verified by hash.

ZeroGPU availability and per-account daily quotas apply. Signing in to Hugging Face can provide the account's free quota. No paid hardware or subscription is requested by this application. Results and downloaded references use the Space's temporary disk and can disappear on restart.

The Windows installation remains a separate alternative, with its own dataset and results. This cloud deployment does not modify the PC's model environment.

The existing checkpoint is used without retraining. The complete 100STYLE dataset remains preserved on the owner's D: drive; this Space exposes only the 47-style subset requested for the demonstration. Optional DDIM content inversion is experimental and may change gestures or trajectory.

Sources and licenses

• Style-SALAD, commit 4faebb83ccda8630e27fbe2189aed1928f4623a0. Authored source is MIT; the released checkpoint has CC BY-NC-ND 4.0 terms and is downloaded unchanged. See upstream/LICENSE-CHECKPOINTS.md.
• SALAD provides the backbone and normalization assets, downloaded from its official public release. Third-party terms remain applicable; see upstream/THIRD_PARTY_NOTICES.md.
• CLIP, downloaded at the pinned official revision 3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268.
• 100STYLE — Ian Mason, Mason, Starke and Komura (2022), CC BY 4.0; HumanML3D features are from the official SMooDi processed release.
• BVH utility attribution and the complete MoMask MIT notice are preserved in adapter/THIRD_PARTY_NOTICES.md and adapter/_bvh/LICENSE.
• ZeroGPU documentation.

No credentials are stored in this repository. Uploads are read with allow_pickle=False; official checkpoints use weights_only=True.