SkinTokens for the browser (ONNX, fp16)

An ONNX export of SkinTokens (VAST-AI-Research, MIT) split into stages so it can run in a web page with onnxruntime-web on WebGPU. It predicts a skeleton and skin weights for a 3D mesh. This is what powers the free in-browser auto-rigger at afk.ai/rig.

file	what it is
`mesh_embed`, `mesh_data`, `mesh_cross`, `mesh_self`	mesh encoder (512 shape tokens)
`vae_embed`, `cond_data`, `cond_cross`, `cond_self`	skin VAE condition encoder
`dec_latent`, `dec_query`	skin VAE decoder (per-bone latent, per-point weights)
`llm_step2`	the 28-layer autoregressive model, one decoding step with a shared-prefix KV cache
`constants.json`	tokenizer and sampling constants the stages need

Weights are fp16 (the Fourier embedding stages stay fp32). Total download is about 1.15 GB.

The original model and code are MIT licensed by VAST-AI-Research; see LICENSE. This export changes only the format.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for burtonda/skintokens-web

Base model

VAST-AI/SkinTokens

Quantized

(2)

this model