SkinToken Download Models

This is not a workflow node, and it's honest about that. SkinToken Download Models is a one-shot setup utility: it pulls the model files the SkinToken Rig node needs from HuggingFace and drops them where ComfyUI expects them. You queue it once on a fresh install, watch it download, and never think about it again. If you're building a workflow to share or looking for something that does cool things inside a graph - this isn't it. The cool node is its sibling, SkinToken Rig.

What it downloads is controlled by three checkboxes, all on by default:

• download_tokenrig - grpo_1400.ckpt, the TokenRig autoregressive model that actually predicts skeletons and skinning weights.
• download_vae - last.ckpt, the FSQ-CVAE that compresses per-vertex skinning weights into discrete tokens (and decodes them back out at inference).
• download_qwen_config - the Qwen3-0.6B config and tokenizer files only. Notice it's just config: the downloader explicitly skips *.bin and *.safetensors, because the Qwen backbone's weights are bundled inside the main checkpoint. You get config.json and friends, not another 1 GB language model.

Together the two checkpoints run to roughly 1.6 GB, and they land in ComfyUI/models/skintoken/ - a folder the pack registers with folder_paths on import, so the Rig node's ckpt_name dropdown sees them. You get two string outputs: status (a "Ready:" list of the exact paths pulled) and models_dir (the folder itself). That's the whole job.

Here's the gotcha worth knowing before you bother: the Rig node auto-downloads anything missing on first run anyway. Its code calls an "ensure required models" step before it loads anything, so if you skip this node entirely, your first rig run just takes longer. This node exists to pre-pull the files - genuinely useful if you know you're about to run a big batch, or you want to confirm your download actually succeeded without burning a rig pass. It also won't travel with a workflow, so it's per-machine setup, not something to wire into a saved graph.

Installation is the pack install, not extra work beyond it: search ComfyUI-SkinToken in ComfyUI Manager, or git clone https://github.com/Rizzlord/ComfyUI-SkinToken into custom_nodes and pip install -r requirements.txt. Then restart, queue the node with all three toggles on, and wait out the download.

Troubleshooting is minimal because the node barely does anything. A failed download (HF hiccup, dropped connection mid-file) just means re-queueing - the Rig node's auto-download will retry anyway. If you see files under models/skintoken/ in your models dir, it worked. One honest caveat: this pack is young and unproven in the wild, and 1.6 GB of model files is the price of entry before you ever run a rig. Cheap enough, but set expectations accordingly.