📥 VAE Download

Small node, narrow job: paste a URL to a VAE file, name it, and get a loaded VAE out - downloaded on the spot if you don't already have it. It's the SDVN download-and-load wrapper for the one component that sits between latent space and actual pixels.

What a VAE is and when you actually need one

The VAE is the codec that converts between the compressed latent space where diffusion happens and the RGB pixels you see. Here's the thing most beginners get backwards: on modern models you usually don't attach a separate VAE, because the checkpoint bakes the right one in. The "load an external VAE" habit comes from the SD 1.5 days, when the fix for washed-out grey colors was to drop in the 840k VAE file. That was additive and everyone did it.

Everything since has moved the other way. Attaching a mismatched VAE to a model that already has a good one baked in is now the mistake - you can make colors worse, not better. So the honest use for this node is specific: you're loading a bare diffusion model (Flux, SD3) that genuinely ships its VAE as a separate file, or you're deliberately swapping a checkpoint's decoder for a known-better one (people do this to undo the over-smoothing some 2026 decoders were trained into). If you're not in one of those cases, you probably don't need this node at all.

The inputs and outputs that matter

• Download_url - a civitai or Hugging Face link to the VAE file. Page address or direct download link both work.
• Url_name - the filename to save as, e.g. flux_vae.safetensors. Rename it from the default model.safetensors so you can find it later.

Output is a single VAE, which wires into your VAE Decode (and VAE Encode, for img2img/inpaint) node.

Installing it

ComfyUI Manager: search SDVN_Comfy_node, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node

Then pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt from the ComfyUI root, and restart. Windows/macOS users need aria2c installed for the download to run.

Where people get burned

Using the wrong VAE for the model. This is the whole trap. The right VAE is the one the checkpoint was trained against, not the one with the best reputation on some forum. A Flux VAE on an SDXL model, or vice versa, produces mangled colors or straight garbage - the latent channel counts don't even match (SDXL is 4-channel, Flux is 16). Match the VAE to the model family.

Attaching one when the model already has it baked in. If your outputs look worse after adding a VAE, try removing it. Modern checkpoints don't need the SD 1.5-era external-VAE ritual.

Death by a thousand round trips. Not this node's fault, but worth knowing: every encode-decode pass through a VAE costs a little quality. Chained img2img and repeated inpaint passes grind an image down over time. That's a workflow-design thing, not a bad-VAE thing.

Download failed. Missing aria2c on Windows/macOS, or a gated repo. Check the console.