Instant Load Dual CLIP

If you've built an SDXL or Flux workflow, you already know DualCLIPLoader - the node that fuses two text encoders into a single CLIP object for conditioning. Instant Load Dual CLIP is that node, restamped on InstantTensor's streaming loader. Same dropdowns, same single CLIP output, same result - minus the extra in-memory copy of both encoders while they load.

Why reach for it: text encoders are the sneaky memory hogs of modern workflows. Flux's T5-XXL alone runs ~11GB in fp16, and a dual load brings up two encoders at once. Peak memory during that load is exactly where small-VRAM setups die before the sampler ever gets to run. Streaming each encoder's tensors straight into the model parameters means you never hold a full state dict on top of the models you're keeping.

The three inputs that matter

Only three, and two of them are just dropdowns:

• clip_name1 / clip_name2 - pick both encoders the recipe needs from everything in ComfyUI/models/text_encoders.
• type - the recipe enum, and this is the input people actually get wrong. It tells the loader what pairing you're building, and the dropdown runs 12 deep: sdxl, sd3, flux, hunyuan_video, hidream, hunyuan_image, hunyuan_video_15, kandinsky5, kandinsky5_image, ltxv, newbie, ace.

The author ships a cheat sheet in the node description, and it's worth repeating:

• sdxl → clip-l + clip-g
• sd3 → clip-l + clip-g, or clip-l + t5, or clip-g + t5
• flux → clip-l + t5
• hidream → t5 + llama (at least one of the two, both recommended)
• hunyuan_image → qwen2.5vl 7b + byt5 small
• newbie → gemma-3-4b-it + jina clip v2

The type you pick must match what the model actually needs - that's how ComfyUI decides which CLIPType to build. Match the recipe and everything downstream is identical to the built-in node.

Output

• CLIP - wire it into CLIPTextEncode. Done.

The gotchas specific to this node

The big one is that there is no CPU device option. ComfyUI's stock DualCLIPLoader lets you load encoders on CPU so they can be swapped in and out of VRAM; this node deliberately strips that option because InstantTensor is CUDA-only. If your card fits both encoders, great. If you were leaning on CPU offload to survive a dual-encoder load, this node is not your friend - the normal built-in node will treat you better there.

The pack-level caveats apply too: .safetensors files get the streaming fast path, everything else falls through to ComfyUI's regular loader, and if streaming can't handle a file it raises instead of doubling memory. And this pack is young - a single commit as of May 2026 - so there's no community troubleshooting lore if something bites.

How to install

Via ComfyUI Manager, search ComfyUI-InstantTensorLoaders and install (accept the dependency install - that's what pulls in instanttensor). Or manually:

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

Restart ComfyUI. The node shows up under advanced/loaders/instanttensor. If you get "The instanttensor package is required for CUDA instant loading," it means the pip dependency didn't land in your environment - install it with pip install instanttensor in the venv that runs ComfyUI, then restart.