Downloads · 30 days
0
ardy-web-onnx
sprited/ardy-web-onnx
ardy-web-onnx is a machine learning model from sprited. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Browser-ready exports of NVIDIA's ARDY autoregressive motion diffusion model, used by the standalone web demo in sprited-ai/ardy (branch mps-support, folder web/). The client downloads these files at load time; nothin…
Access
Public
Updated Sep 3, 2026
Repo size
6.6 GB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.data5 GB · 75%
At a glance
- License
- other
- Access
- Public
- Created
- Sep 2, 2026
- Updated
- Sep 3, 2026
- SHA
- 02bb5dcd
Base models
Model card
From the Hugging Face model README
ARDY for the browser (ONNX, WebGPU)
Browser-ready exports of NVIDIA's ARDY autoregressive motion
diffusion model, used by the standalone web demo in sprited-ai/ardy (branch
mps-support, folder web/). The client downloads these files at load time; nothing runs on a server.
| folder | base model | contents |
|---|---|---|
core-rp-20fps-h40/ | nvidia/ARDY-Core-RP-20FPS-Horizon40 | window_cond.onnx (779 MB, fp32; used by the app) takes the denoiser's text conditions cond [1,1,2048] (root 1024 + body 1024, text_proj.onnx = the two embed_text layers maps the 4096-d LLM2Vec feature to it; distilled small encoders such as intsuc's MiniLM student produce it directly). window.onnx (813 MB) is the same graph with the text_feat [1,1,4096] input. Each is one autoregressive window — 4 history frames + 40 generated frames at 20 fps, 10 denoising steps with classifier-free guidance, decoder and forward kinematics — as a single graph. Inputs: history [1,4,330], text_feat [1,1,4096], noise [1,10,148], cfg_weight_text [1], cfg_weight_cstr [1]. Outputs: normalized motion features [1,44,330], joint positions [1,44,27,3], joint rotations [1,44,27,3,3]. Plus window.json (shapes, fps, bootstrap history), skeleton.json, skin_cskel27.{json,bin} (Core skin for linear blend skinning), prompts.{json,bin} (20 precomputed prompt embeddings; free-text prompts use text-encoder/). |
text-encoder/ | meta-llama/Meta-Llama-3-8B-Instruct + McGill-NLP LLM2Vec adapters (Built with Meta Llama 3) | ARDY's prompt encoder as INT4 MatMulNBits ONNX in three parts (part_a/b/c, 2.0 + 1.5 + 1.5 GB) so it runs in onnxruntime-web (one worker per part); encoder.json describes the chain, tokenizer/ the Llama-3 tokenizer. See text-encoder/README.md, NOTICE and LICENSES/ (Meta Llama 3 Community License, LLM2Vec MIT). |
Exported with web/export_web_onnx.py from ardy/exports/web.py (WebWindow), which reproduces Ardy.autoregressive_step
bit-for-bit for the text-only case. onnxruntime matches PyTorch to 3e-4 on the motion features.
License
The weights are Derivative Models of NVIDIA ARDY and are distributed under the
NVIDIA Open Model License
(see LICENSE and NOTICE). Licensed by NVIDIA Corporation under the NVIDIA Open Model License.
- License
- other
- Created
- Sep 2, 2026
- Updated
- Sep 3, 2026