VAE Loader

It's the most boring node in the pack and that's fine - you'll still want it. VAE Loader [Eclipse] loads one external VAE from your models/vae/ folder and hands you the decoded image. Its one genuinely useful extra is that it outputs the filename too (vae_name), which comes in handy when you're stamping metadata or naming saved files.

You reach for it in two situations. First, when your model file has no baked VAE - UNet, Nunchaku, and GGUF flows all need an external one, and this is the pack's answer to that. Second, when you want to override a baked VAE. This is a real, well-worn trick in the community: many people swap the stock checkpoint VAE for a fine-tuned one (like the classic SDXL VAE fix) because the baked version produces slightly duller, flatter colors. If your images look washed out and no amount of prompt tweaking helps, a better VAE is often the actual fix.

Inputs

• vae_name - the file. Straightforward.
• disable_offload - default true, meaning keep the VAE on the GPU. Leave it on unless VRAM is genuinely tight: VAE decode is one of those chunky, memory-hungry steps, and offloading it means shuffling tensors back and forth for little gain at typical image sizes. Toggle it off only when big-image decode is OOM-ing.

Outputs

• vae → VAE Decode (and VAE Encode if you're doing img2img here).
• vae_name → the filename as a STRING, for your Save Image metadata or a text display node.

Mechanically it mirrors ComfyUI's stock VAELoader under the hood (comfy.sd.VAE), with one improvement the pack is upfront about: enhanced Wan 2.1 architecture detection. Wan 2.1's VAE files have tripped up plain loaders in the past, so if you're building a Wan video workflow, this is the VAE loader that actually recognizes what it's looking at.

Installing

Same as every node in the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_SmartModelLoader.git
cd ComfyUI_SmartModelLoader
python -m pip install -r requirements.txt

or search ComfyUI Smart Model Loader in ComfyUI Manager, then restart.

Where people trip

• No VAE selected / None - pick one from the dropdown; unlike the Smart Model Loader there's no baked fallback here, this node only loads external files.
• Blurry or pixelated decode - that's not the node, that's a VAE/model mismatch. Make sure the VAE's latent channels match your model (SDXL vs SD1.5 VAEs are not interchangeable).
• OOM on a huge image - flip disable_offload off, or use tiled decoding in the sampler node instead of fighting it here.