Disable VAE Offload (VAE Utils)

If your decode step is crawling and you've narrowed it down to "ComfyUI keeps shuffling the VAE around," this is the node you're looking for. It's a tiny wrapper: a VAE goes in, the same VAE comes out, and in between it flips off ComfyUI's habit of parking the decoder in system RAM between uses. That's the whole job. Not glamorous, occasionally exactly what you need.

Why you'd reach for it

ComfyUI is aggressive about VRAM. By default it treats the VAE like any other model it can evict - after a decode finishes, the VAE gets pushed off the GPU to free room for the next thing, then dragged back onto the GPU the next time you decode. For a single image that shuffle is invisible. For anything that decodes over and over, it stops being invisible.

That "over and over" is the Wan story, which is the ecosystem this whole pack lives in. Video means many frames, tiled decoding means many tiles, batch work means many passes - and if the VAE is bouncing on and off the GPU for each one, you're paying a load penalty you never asked for. Keeping the decoder resident on the card means it's just there, ready, every time. The trade is honest and it's the only trade: you spend the VRAM the VAE occupies and you get the reload time back. On a card that's already tight, forcing the VAE to stay resident can be the thing that tips you into an out-of-memory error, so this is a lever for people with headroom, not a free win for everyone.

How it works

There's no magic here and that's the point. The node takes your VAE and marks it so ComfyUI's model manager leaves it on the GPU instead of offloading it. Everything downstream - your decode node, your tiled decode, whatever - then uses a decoder that never has to be re-fetched from RAM. It changes a scheduling decision, not the math; your pixels come out identical, they just come out sooner.

Inputs and outputs

Two inputs, one output, nothing to overthink:

• vae - the VAE you want to pin. Wire your VAE loader straight into it.
• disable_offload - a boolean, and it defaults to true. So dropping the node in with defaults already does the thing. Set it to false and the node becomes a no-op pass-through, which is handy if you want to A/B the behavior from a single toggle without rewiring.

The output is a VAE, so it slots in anywhere a VAE goes. The intended shape is: VAE loader → Disable VAE Offload → your decode node. It sits mid-cable and quietly changes the residency policy of whatever passes through.

Installing it

The pack has no dependencies at all - the author's own words are that there's nothing needed "outside of what ComfyUI already uses." So it's about as easy as installs get. Through ComfyUI Manager, search for ComfyUI-VAE-Utils and install. Or clone it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/spacepxl/ComfyUI-VAE-Utils

Restart ComfyUI afterward. No model download is required for this node specifically - the only file the pack ever asks you to fetch is for its Latent Upscale node, which is unrelated to this one.

Common issues

The failure mode to watch for is the obvious one: you pin a large VAE on a card that was relying on that offload to survive, and now you OOM where you didn't before. If enabling this pushes you over the edge, that's your answer - you don't have the spare VRAM for it, turn it back off. It's a speed-versus-memory dial, not a correctness fix, so if a decode is producing wrong output rather than slow output, this node won't touch that; you're looking at a VAE mismatch or a normalization problem instead. And keep expectations calibrated to what you're decoding - pinning the VAE only pays off when you decode repeatedly. For a one-shot single image, you won't notice a thing.

For context on who's behind this: spacepxl is a low-level Wan-tooling person - the Wan control LoRAs and a chunk of the video-training research others build on - so a pack of small, sharp utilities like this is exactly on-brand. Niche, but the niche is real.