VAE Downloader

Of the four ModelUtils/Downloader nodes in this pack, this is the one you'll probably run the least often, and for a good reason: VAEs are rare. There have only ever been about 200 of them published on CivitAI total, against three-quarters of a million LoRAs - VAEs are infrastructure, shipped by labs, not the kind of thing the community churns out variants of. So most people don't have a folder full of mystery VAE files the way they do checkpoints or LoRAs. But if you're one of the people who's started keeping a small stash on purpose - the SD 1.5 840k fix, an SDXL fp16-fix build, maybe a Wan 2.1 VAE you grabbed specifically to fight Qwen-Image's over-smoothing - this node fetches metadata for all of them in one pass, feeding VAEInfoLoader so you can actually tell them apart later.

How it works. It scans a folder, works out which files it can identify, and pulls down whatever metadata, example images, and workflows exist for them from a hosted model repository - the nsfw_level filter and the api_key requirement both point at something CivitAI-shaped, matching files by content hash rather than filename.

Inputs that matter. subdirectory (default /) is the folder to scan, relative to your VAE directory. recursive (default on) controls whether it walks into subfolders. nsfw_level (default "All") filters results by the content rating on the source - not that VAE listings tend to be NSFW-flagged much, but the option's there and shared across all four downloaders in this pack. max_examples (0–100) is how many example images/workflows to grab per file. threads (1–32, default 4) controls parallelism. api_key is your key for the source repository, needed for anything gated behind it.

Output. status (STRING), a summary string. This is an output node, so it runs as a terminal step in the graph regardless of whether you connect status to anything - same behavior as SaveImage.

Installing it. Through ComfyUI Manager - search "Model Utility Toolkit" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils

then restart ComfyUI.

Troubleshooting. The most likely outcome of running this against a typical VAE folder is: not much comes back. A lot of the VAE files people actually use are bare tensor extracts pulled straight from a HuggingFace repo config rather than something with a CivitAI-style listing - vae-ft-mse-840000-ema-pruned and the FLUX or Wan autoencoders are official artifacts, not community uploads, and this node has nothing to match them against on a site built around community model pages. Don't read a blank status as broken; for VAEs specifically, there's often genuinely nothing on the other end. Where it will help is the handful of community-uploaded VAE variants that do have listings - run it, then use VAEInfoLoader to check what actually got cached.