QwenVL HuggingFace Downloader 📥

Here's the thing about this node: it has almost nothing to do with Qwen-VL. It's a plain file fetcher that happens to ship inside the ComfyUI-QwenVL pack, and it exists because of a small lie in that pack's pitch. The QwenVL nodes "automatically download models on first use" - and they do, for the Transformers models. But the GGUF route is explicitly marked manual download in the README, and that's where this node earns its keep. Run it once, and a multi-gigabyte GGUF lands in the right folder while you go make coffee, no wget and no dragging files around.

What it actually does

Under the hood it's a thin wrapper over huggingface_hub - hf_hub_download for a single file, snapshot_download for a whole repo. You give it a HuggingFace repo ID, it pulls the weights into ComfyUI/models/, and it hands back the local path. Nothing loads a model, nothing touches your GPU. It's plumbing, and it's honest plumbing.

Files land in a neat ComfyUI/models/<folder>/<author>/<repo> layout, so the pack's GGUF nodes - which default to LLM/GGUF - find what you downloaded right away. The snapshot_download path also skips *.msgpack, *.h5, and coreml/*, which quietly saves you a few hundred MB of junk on the bigger repos.

The inputs that matter

• repo_id - the HuggingFace repo, like Qwen/Qwen3-VL-4B-Instruct-GGUF. This is the one you'll always change.
• filename - a specific file to grab, e.g. Qwen3VL-4B-Instruct-Q4_K_M.gguf. Leave it empty and it downloads the entire repository, which can be tens of gigabytes. Empty is a choice; make it deliberately.
• save_folder - where to put it: LLM/GGUF, LLM/hf, checkpoints, or loras. Default is LLM/GGUF, and that's the one that matches the GGUF nodes.

Output is a single downloaded_path string - the local path to the file or repo folder. It's not a model object you can plug into a loader; it's a path you can read, log, or feed into any node that accepts a file string. One detail that tells you a lot: if the download fails, it doesn't raise an error - it returns the error text as the output string. So check that output when things seem to silently go nowhere.

Install

Same as the whole pack - ComfyUI Manager (search ComfyUI-QwenVL) or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-QwenVL.git
cd ComfyUI-QwenVL
pip install -r requirements.txt

Then restart ComfyUI. The downloader itself only needs huggingface-hub (and hf_xet for the fast Xet transport), but the pack's requirements drag in the whole transformers stack - torch, transformers, bitsandbytes, accelerate, opencv-python. If you only want the downloader you can install just huggingface-hub and skip the rest, though the other nodes won't work until you finish the job.

Gotchas

• Don't use it for the plain QwenVL nodes. Those auto-download to ComfyUI/models/LLM/Qwen-VL on first run anyway. Pre-downloading into the downloader's LLM/hf folder won't help them - that's not where they look. This node is for the GGUF path and for pulling arbitrary HF assets.
• Gated repos fail quietly. The node uses your normal HuggingFace credentials, so if a repo needs a login, run huggingface-cli login in your ComfyUI Python environment first, or the "download failed" string will be all you get.
• Disk space. A full repo snapshot is huge, and snapshot_download doesn't resize anything. Check free space before you hit "run."
• First-run surprise: the pack's security posture is worth a beat. An LLM/VLM node is expected to reach the network, which is exactly the shape of thing that's been weaponized once before in this ecosystem (ComfyUI_LLMVISION). 1038lab's pack is open and GPL-3.0 - but the habit of reading what a fresh node downloads before running it is the right one, and this node makes it easy: you can see exactly what repo and file it's pulling.