TREEIndustries/ARDY-Core-RP-20FPS-Horizon8-ONNX

ARDY Core RP 20 FPS Horizon8 ONNX

Validated ONNX conversion of nvidia/ARDY-Core-RP-20FPS-Horizon8, released by TREE Industries for responsive real-time motion generation.

Model

Property	Value
Skeleton	ARDY Core, 27 joints
Motion rate	20 FPS
Generation horizon	8 frames
Frames per token	4
Maximum runtime window	64 tokens / 256 frames
Latent dimension	128
Text feature dimension	4096
Diffusion steps	10
Source revision	257a0843a10bf5201065963d7bca2791e9393a7a

Three graphs form the runtime model:

• encoder.onnx: normalized motion history to FSQ latent tokens
• denoiser.onnx: guided, text-conditioned latent motion denoising
• decoder.onnx: latent tokens to local-root and body motion features

model_contract.json records the complete skeleton and tensor contract.

Builds

• fp32/ contains exact FP32 reference exports.
• directml-mixed-fp16/ contains compact runtime builds with FP32 graph I/O and sensitive operations plus selected FP16 constant-weight matrix islands.

The denoiser consumes a 4096-value text feature. Raw runtime prompts can be encoded with TREEIndustries/Llama-3-ARDY-Text-Encoder-ONNX.

Validation

All six graphs passed onnx.checker. The mixed-FP16 graphs were executed against FP32 with identical deterministic 64-token/256-frame inputs using ONNX Runtime 1.20.1 and DirectML. Shapes matched and every output was finite. Measured differences are in VALIDATION.md.

No TREE plugin code, Unreal Engine assets, or Blueprints are included.

License

Distributed under the NVIDIA Open Model License.

Licensed by NVIDIA Corporation under the NVIDIA Open Model License

See NOTICE and THIRD_PARTY_NOTICES.md.

Citation

@article{zhao2026ardy,
  title     = {ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation},
  author    = {Zhao, Kaifeng and Petrovich, Mathis and Zhang, Haotian and Wang, Tingwu and Tang, Siyu and Rempe, Davis},
  journal   = {ACM Transactions on Graphics (TOG)},
  year      = {2026},
  volume    = {45},
  number    = {4},
  articleno = {86},
  doi       = {10.1145/3811284}
}