Hive 模型下载器/Model Downloader - Github:﹫luguoli

Let's be honest about what this node is: a download manager wearing a node costume. You paste a model file's URL into it, pick a folder, hit the button, and the file lands in the right place on your disk. No terminal, no digging through your ComfyUI install trying to remember where checkpoints go. If the gap between "I found a great model" and "I'm actually generating with it" has always been the file-management part, this is aimed straight at you.

It's one of two utility nodes in ComfyUI-Hive, a bigger community/social pack - chat rooms, workflow sharing, prompt tools - from a solo author who publishes as @luguoli (the pack is GPL-3.0, and its reddit footprint is a single low-engagement self-post, so this is a small, one-person project, not a household name). The downloader exists to serve that pack's pitch, which the README states plainly: "Can't access HuggingFace? Don't know how to download models? Don't worry! Built-in mirror addresses..." The real target user is someone on a network where HuggingFace or Civitai is slow or blocked - paste a mirror link and the node does the tedious part.

How it works

On run, the node finds your ComfyUI install by walking up the directory tree until it hits a folder named models. Then it makes a quick HEAD request to the URL to learn the file size and whether the server supports byte ranges. If it does, the file is split into 4–8 chunks and pulled down in parallel threads - that's the "multi-threaded download" the description promises, each chunk getting its own connection so they don't fight. If not, it falls back to a single stream: slower, but still works. It also checks whether the file already exists and skips the download rather than wasting your bandwidth twice.

The two inputs that matter

• url - the direct HTTP/HTTPS link to the model file. This is the whole job; paste something wrong here and nothing else matters.
• save_directory - a dropdown built live from the subfolders that actually exist under your models directory. Pick where the file belongs: checkpoints, loras, vae, upscale_models, whatever your setup has.

There are no outputs. It's an output node in the ComfyUI sense - it reports its status and progress as text in the node UI and the console rather than passing anything down a wire. The pack's frontend adds a "Start Download" button right on the node, so you trigger it by clicking that.

Installing it

The pack installs like any other custom node:

cd ComfyUI/custom_nodes
git clone https://github.com/luguoli/ComfyUI-Hive

or search ComfyUI-Hive in ComfyUI Manager and click install. There's also a Gitee mirror for mainland China. Its real dependencies are just requests and tqdm - tiny, and pip-installed automatically. You don't need to download any model files to use the pack itself.

The gotchas

• Restart ComfyUI after the download finishes. The README says it, and it's the #1 "I downloaded it but can't find it" cause: ComfyUI scans the models folders at startup, so a fresh file won't show up in any loader until you restart.
• The filename comes from the URL. Links with query strings or no obvious extension can come down as a generic downloaded_model.bin. Rename it to something sensible, with the right extension (.safetensors for checkpoints and LoRAs), or you'll be squinting at a mystery file later.
• Pick the right subfolder. The node only fetches the file; it doesn't register it anywhere. Drop a checkpoint into loras and it simply won't appear in your checkpoint loader.
• Range support varies by host. The multi-threaded speedup only kicks in when the server advertises byte-range support. Some CDNs don't, and the node quietly goes single-threaded.
• Download hygiene still applies. A model file is code that runs when you load it. Grab things from the model author's page or a mirror you trust - this node makes fetching convenient, but it can't make a sketchy source safe.

It won't replace ComfyUI Manager's model browser for everyone. But for the beginner who's never touched a terminal, and the person stuck behind a slow mirror, it's genuinely the least-friction path from link to model that lives inside the graph.