Diffusers Vae Loader

Same job as ComfyUI's native VAE loader, different type system. This node loads a VAE and hands it out as an AUTOENCODER - the type Diffusers Model Makeup expects, not the VAE type native ComfyUI nodes use. If you're not already deep into this pack's Diffusers or StreamDiffusion nodes, you don't need this one; use the built-in VAE Loader instead. If you are, this is how you override the VAE that came bundled with your checkpoint.

Why override it at all? Because a mismatched or missing VAE is the single most common "why does my output look grey and washed out" complaint in the whole ecosystem - the model can technically still decode the latent without the right VAE, it just does it badly. The fix is always the same regardless of which loader you're using: load the VAE the checkpoint was actually trained against, not whichever one has the best reputation.

The inputs and outputs that matter

There's one input - vae_name, a dropdown populated from your VAE folder - and one output, AUTOENCODER, which feeds into Diffusers Model Makeup alongside a pipeline and scheduler. That's the entire node. It doesn't decode anything itself; it just loads the file and exposes it in the type the rest of this pack's plumbing expects.

How to install it

ComfyUI Manager → Install Custom Nodes → search ComfyUI-Diffusers, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Limitex/ComfyUI-Diffusers.git
cd ComfyUI-Diffusers
pip install -r requirements.txt

Restart ComfyUI afterward. No extra weights ship with the pack - you supply your own VAE files, same as always. If you're planning to go all the way to the StreamDiffusion nodes, the README also has you clone the StreamDiffusion repo and run its TensorRT installer, but that's not required just to use this loader.

Common issues & troubleshooting

"Missing node type: DiffusersVaeLoader." This is a documented real-world stumble - someone loaded a shared workflow, got exactly this error with zero context on which pack it came from, and had to search the web to land on Limitex/ComfyUI-Diffusers before they could fix it. If you hit this error on any node starting with Diffusers, this pack is almost always the answer. Install it and restart.

Empty dropdown. The loader reads whatever's sitting in your VAE folder - if nothing shows up, confirm the file actually landed there and that you restarted ComfyUI rather than just refreshing the browser tab.

Output still looks grey or washed out after loading a VAE here. Two likely culprits: you loaded the VAE but never actually wired its AUTOENCODER output into Diffusers Model Makeup (a disconnected VAE loader does nothing), or the VAE doesn't match the checkpoint's architecture - an SD 1.5 VAE on an SDXL-family checkpoint, for instance, produces a poor decode rather than an error. Match the VAE to what the checkpoint was trained on, not the loudest-recommended file online.

It loads fine but nothing downstream accepts it. AUTOENCODER is a diffusers-pack-specific type. It only connects to nodes inside this pack's Diffusers/StreamDiffusion family - it will not plug into ComfyUI's native VAE Decode or VAE Encode nodes, and vice versa. If your graph is otherwise all-native, you probably want the built-in VAE Loader, not this one.