Instant Load CLIP

Instant Load CLIP is the pack's take on ComfyUI's stock CLIPLoader - the node that loads a single text encoder when you don't need a full checkpoint's worth of parts. Same job, same dropdowns, one output, and for .safetensors files the weights stream straight into the model instead of sitting in a temporary state dict first. On a big T5-XXL that difference is real: it's the difference between the encoder fitting next to your UNet and an OOM right at load time.

This is the node you reach for with single-encoder architectures: SD 1.5's clip-l, Stable Cascade's clip-g, SD3 and Mochi's t5-xxl, Wan's umt5-xxl, and a bunch of the newer recipe-based models that ship one encoder at a time.

The two inputs that matter

• clip_name - dropdown over everything in ComfyUI/models/text_encoders.
• type - the recipe enum. The dropdown is 25 entries deep, and picking the wrong one is the number one failure mode here, because the type decides how ComfyUI builds the encoder.

The author packs a recipe sheet into the node description, and it's genuinely useful:

• sd → clip-l
• stable cascade → clip-g
• sd3 → t5 xxl / clip-g / clip-l
• stable audio → t5 base
• mochi → t5 xxl
• cogvideox → t5 xxl (226-token padding)
• cosmos → old t5 xxl
• lumina2 → gemma 2 2B
• wan → umt5 xxl
• hidream → llama-3.1 (recommended) or t5
• omnigen2 → qwen vl 2.5 3B
• lens → gpt-oss-20b
• pixeldit → gemma 2 2B elm

If you load clip-l and tell it stable_cascade, you'll get garbage conditioning or an outright error. Match the file to the recipe and it just works.

Output

• CLIP - feed it to CLIPTextEncode, same as any other CLIP output.

Gotchas

Two things worth knowing before you build a workflow around this node. First, there is no CPU device option - ComfyUI's built-in CLIPLoader can load the encoder on CPU and swap it in and out of VRAM, and this node deliberately hides that option because InstantTensor is CUDA-only. If CPU offloading was your survival strategy, the stock node stays your friend.

Second, the pack is young and niche (a single commit as of May 2026, no community lore to speak of). .safetensors files get the streaming fast path; other formats fall through to ComfyUI's normal loader; and if streaming can't handle a file, the node raises rather than doubling memory. Read that as intended behavior, not breakage.

How to install

ComfyUI Manager → search ComfyUI-InstantTensorLoaders → install and accept the dependency install (that's the instanttensor pip package, the pack's only requirement). Or manually:

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

Restart ComfyUI, and the node appears under advanced/loaders/instanttensor. If you hit "The instanttensor package is required for CUDA instant loading," the pip package didn't reach your environment - pip install instanttensor in the venv ComfyUI runs from, then restart.