ardy

What it is

ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation. It is an autoregressive diffusion model designed for interactive motion generation with online text prompts and flexible long-horizon kinematic constraints (root paths/waypoints, full-body keyframes, and sparse joint positions/rotations) with real-time responsiveness. The repo provides code, checkpoints, and demos for pre-trained ARDY models introduced in the SIGGRAPH paper.

How it works

The project implements an autoregressive generation loop that encodes text prompts into motion representations, optionally applies kinematic constraints, and decodes to joint positions. It includes an interactive demo that streams text-to-motion, supports waypoint and velocity controls, and allows sampling constraints from a motion file. The text encoder relies on a gated model and requires a Hugging Face token at runtime. Model loading and generation are exposed through scripts and mixins in the codebase (e.g., load_model, autoregressive_step, _generate_step).

Getting started

Setup instructions assume Ubuntu 22.04 with RTX 4090 and Python 3.11. In a Python 3.10+ environment, install PyTorch matching CUDA, then install ARDY in editable mode with all features:

conda create -n ardy python=3.11 -y
conda activate ardy
# Install PyTorch for your CUDA version first — see https://pytorch.org/get-started/locally/
# For example:
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
pip install -e ".[all]"

The README includes details about build requirements (CMake >= 3.15, C++17 compiler) and optional install groups: core, demo, and trt. It also explains setting up a Hugging Face token via:

hf auth login

or placing the token in ~/.cache/huggingface/token.

Checkpoints are available and will be downloaded automatically when used in the demo. The README lists several ARDY model variants with their FPS and horizon.

Launching the interactive demo:

python scripts/run_demo.py

To run the text encoder server in the background:

python scripts/run_text_encoder_server.py

Recent releases

• The repository shows latest releases as none under the provided entry "RELEASES (latest 0): - none"; no explicit versioned release notes are listed in the provided content.

Traction

195 stars on the repository as of the provided data.

Behind the repo

Not applicable here as there is no linked startup/company section in the provided facts.

Caveats

• License: Apache-2.0
• Created: 2026-07-08; Last push: 2026-07-10
• The text encoder requires Hugging Face access and a token, which must be provided at runtime; instructions are included in the README. Checkpoints are downloaded automatically when used in the demo, so manual download is not required.