Instant Load VAE

The VAE is the smallest major component in a ComfyUI graph - and the one that decides whether your image actually appears at the end. It encodes your prompt's latents and, more painfully, decodes them back into pixels, which is where low-VRAM setups often fall over on big resolutions. Instant Load VAE is the pack's drop-in for ComfyUI's VAELoader: one dropdown, one VAE output, same models/vae folder. For .safetensors files the weights stream straight into the decoder's parameters instead of building a full state dict first.

Honest framing: the VAE is small enough that the memory win here is smaller than the checkpoint or UNet nodes deliver. You're shaving a load spike, not unlocking a model that wouldn't fit. But the node is a one-to-one replacement - same dropdown, same output - so if you're already running the pack for the big loaders, there's no reason not to swap this one too. Consistency is a feature.

The one input and output

• vae_name - dropdown of everything in ComfyUI/models/vae. The dropdown also lists ComfyUI's built-in special entries - the approximate TAESD VAEs and, on newer ComfyUI, pixel_space - and those intentionally fall through to ComfyUI's normal loader rather than the streaming path. Only actual .safetensors VAE files get the InstantTensor treatment.
• VAE - wire it into VAE Decode (and into VAE Encode if you're doing img2img or inpainting).

Common issues

The pack rules, one more time: CUDA only, so a CPU-torch ComfyUI errors immediately; the instanttensor pip package must be in the environment or you get "The instanttensor package is required for CUDA instant loading"; and non-safetensors files (.pt VAEs, TAESD approximations, pixel_space) use the stock loader, which is fine - those aren't what this node is optimizing. If a workflow fails with this node, verify the dependency first: it's the pack's single requirement and the most common thing to be missing.

How to install

ComfyUI Manager → search ComfyUI-InstantTensorLoaders → install, accepting the dependency install (that pulls in instanttensor). Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/redstonewhite/ComfyUI-InstantTensorLoaders

Restart ComfyUI; the node is under loaders/instanttensor. If the dependency didn't land in your environment, pip install instanttensor in the venv ComfyUI runs from and restart. And the standing caveat: this is a young pack (single commit, May 2026) with no community track record - test it against the stock VAELoader and keep whichever behaves for your setup.