Instant Load Checkpoint

Ever watched a big checkpoint load and seen your VRAM meter spike to a number that makes no sense? That's ComfyUI's normal loading behavior: for a moment you hold a full state dict and the model it's about to become. Instant Load Checkpoint is a drop-in swap for the stock Load Checkpoint node that streams the weights straight into the model's parameters instead, so you only ever hold one copy.

This is the pack's flagship node, and if you only install one of these Instant loaders, it's the one. Same input, same three outputs, same models/checkpoints folder - it slots into the standard txt2img template exactly where the built-in node sits. The only thing that changes is the memory profile under the hood.

How it works

The trick is InstantTensor, a CUDA library that the pack builds on. For .safetensors files the node doesn't actually load the file first. It reads the safetensors header, builds a skeleton state dict of "meta" tensors - shape and dtype info, zero actual bytes - lets ComfyUI detect the architecture and construct the model from that, then streams each tensor from InstantTensor's reusable transfer buffer directly into the model's parameters with a copy_(). The one model-sized allocation left in VRAM is the model itself.

That's the whole pitch: no full CUDA state dict sitting next to the model. The author calls the naive alternative the "Spark double-memory problem," and this node exists to kill it. When the model finishes loading it's handed to ComfyUI's model manager, so unloading and swapping between checkpoints still work the way you expect.

The inputs and outputs

There's exactly one input, which is the nice part for a beginner:

• ckpt_name - a dropdown of everything in ComfyUI/models/checkpoints. Same list as the built-in loader.

Three outputs, and they wire up exactly like the stock node:

• MODEL - the denoising model; feeds your KSampler.
• CLIP - the text encoder; feeds CLIPTextEncode.
• VAE - decodes latents to pixels; feeds VAE Decode.

Honest caveats

It only takes the fast streaming path for .safetensors (and .sft). A legacy .ckpt falls through to ComfyUI's normal loader - still works, just no memory magic. And note that if the streaming path hits a file it can't handle, it raises an error rather than silently falling back, because a fallback would recreate the double-memory problem. That's not a bug; that's the node refusing to betray its purpose. If you see it, use the built-in loader for that file.

Two environment gotchas, and they're the ones that actually bite:

• CUDA only. The node has no CPU path by design. Running ComfyUI with CPU-only torch means an immediate error, not a slow fallback.
• The instanttensor pip package must be installed. This is the pack's single dependency (see requirements.txt). If it's missing you get "The instanttensor package is required for CUDA instant loading."

One more thing worth knowing: this pack is brand new - a single commit as of May 2026, zero community track record. It works, but there's no lore to Google if something odd happens. If you share workflows that use it, everyone who opens them needs the pack installed or they'll get a missing-node error.

How to install

Through ComfyUI Manager, search for ComfyUI-InstantTensorLoaders and install. Or manually:

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

Restart ComfyUI afterward. When Manager offers to install dependencies, accept - that's what pulls in instanttensor. If your ComfyUI environment is managed outside Manager's venv, install it yourself with pip install instanttensor in the same environment that runs ComfyUI. Then reload, and the node appears under the loaders/instanttensor category.