GGUF VAE Loader

The pitch for this node, straight from the author's own README, is basically "you can't imagine it works, but it really works" - and the pitch is about convenience, not VRAM savings. GGUF VAE Loader loads a VAE file whether it's a .gguf or a regular .safetensors, so you stop needing two different loader nodes depending on which format you happened to download.

Why you'd reach for it

A VAE is small to begin with - a couple hundred megabytes, nothing like the multi-gigabyte checkpoint it decodes for. Quantizing it isn't really where the VRAM win is. The actual value here is that this one node reads either file type transparently, which matters more than it sounds once you're mixing gguf checkpoints (that sometimes ship with a matching gguf VAE) and safetensors VAEs (the far more common case) in the same workflow - you don't have to remember which loader goes with which file.

How it works

Point it at a VAE file and it outputs a VAE object the same way the standard ComfyUI VAE loader does - the difference is purely in what file formats it'll accept going in. Under the hood it's reading the pack's own gguf-connector engine to decode a quantized VAE if that's what you gave it, or just loading the safetensors normally if that's what you gave it instead.

Inputs and outputs

One input: vae_name, a dropdown pulled from your VAE files. One output: VAE, which wires into VAE Decode (to turn latents into images) or VAE Encode (the reverse) exactly like any other VAE loader's output.

Installing it

Via ComfyUI Manager, search gguf and install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/calcuis/gguf

Restart ComfyUI afterward. No extra pip installs needed on the current version - the decoding engine ships bundled with the node. Drop your VAE files, gguf or safetensors, into ComfyUI/models/vae; that's the one folder both formats share here.

Worth knowing: this is a different pack from city96's much more widely used ComfyUI-GGUF. The two are unrelated codebases that happen to solve the same problem and coexist fine side by side - but they're not the same install, so don't expect a tutorial written for one to point you at nodes from the other.

Common issues

Black or broken output after converting a VAE to gguf yourself. This is the one real gotcha the author flags directly: some tensors - VAE weights especially - need to stay at fp32 to work correctly. Cutting or quantizing a VAE more aggressively than it can tolerate is a known way to get a black-screen result. If you hit that, go back to a fp32 or safetensors version of the VAE rather than fighting the quantized one; VAE is small enough that there's rarely a real reason to squeeze it down in the first place.

Empty dropdown. File's in the wrong folder, or wasn't there when ComfyUI started - restart after adding new VAE files, this node reads the folder at launch.

Output looks subtly wrong (color shifts, artifacts). Before blaming this loader, check you're pairing the VAE with a compatible checkpoint family - a mismatched VAE (say, an SDXL VAE on a Flux latent) will produce broken-looking output regardless of which loader you used to bring it in. That's a pairing mistake, not a bug in the node.

If you're not sure whether you even need a separate VAE for your setup: most modern architectures ship the VAE bundled differently than SD1.5-era checkpoints did, so check your model's card before assuming you need to load one manually here at all.