Load VAE (Model Manager)

A VAE is the encoder/decoder half of a diffusion model - the part that turns your latent noise into an actual image on the way out of the sampler. Most checkpoints ship with one baked in, which is why you can ignore this concept for a long time. But plenty of people swap in a dedicated VAE anyway (some community VAEs fix color fringing and produce crisper detail), and if your VAE files live on a Model Manager server instead of your ComfyUI/models/vae folder, Load VAE (Model Manager) is how you get one into the graph.

What it does

One input, vae_name, an enum of models from the server's VAE category (modelId@versionId:name format, (not connected) until you authenticate). On run it downloads the file with a progress bar, caches it in ComfyUI/models/model_manager_cache/vae/, loads it with ComfyUI's VAE constructor, and returns it on a single output: vae. That's it - one socket, one job.

Wire that vae into your VAE Decode node at the end of the workflow. If you're using it deliberately, also make sure your checkpoint isn't feeding its own VAE into the decode - you'd be silently overriding yourself. The common pattern is: Load Checkpoint (Model Manager) for model/clip, this node for the VAE, both from the server, and a clean split of responsibilities.

The whole family shares one connection and one cache. The IS_CHANGED hook keys off the connection version, so the dropdown refreshes when the server changes its model list, and first-load downloads stream with a progress bar.

Install

ComfyUI Manager (search comfyui-model-manager-nodes) or:

cd ComfyUI/custom_nodes
git clone https://github.com/blackpaw-studio/comfyui-model-manager-nodes
cd comfyui-model-manager-nodes
pip install -r requirements.txt

Restart, then hit Connect to Model Manager on the node with your API URL and key (persisted to config.yaml; MODEL_MANAGER_API_URL / MODEL_MANAGER_API_KEY env vars override the file). Dependencies are the pack's usual light pair, requests and pyyaml.

Troubleshooting

• (not connected) - connect first, or your key's wrong (401 kills the list).
• (no models found) - nothing in the server's VAE bucket yet.
• Blurry or washed-out output after swapping VAEs. That's not this node misbehaving - a VAE mismatch (SD 1.5 VAE on an SDXL checkpoint, for example) is the classic cause. Check the base model matches what the server lists.

Most workflows will never touch this node, because the checkpoint loader already hands you a VAE for free. Reach for it when you want the better decoder, and it slots in with no drama.