Load CLIP (Quadruple)

Four text encoders in one node is not a typo - it's what HiDream actually asks for. QuadrupleCLIPLoaderMultiGPU is ComfyUI-MultiGPU's registration of the stock four-slot CLIP loader, grouped with the pack's other loader nodes under the "multigpu" category so you're not digging through the regular node list to find it.

Why four encoders

This is the most conditioning horsepower you'll see thrown at a single model in current ComfyUI. The node's own tooltip gives the exact HiDream recipe: long clip-l, long clip-g, t5xxl, and llama_8b_3.1_instruct - two CLIP variants inherited from the SDXL era, a T5-XXL for language understanding, and an 8B instruction-tuned Llama for good measure. That's not a small ask: an 8B language model alone is comparable in size to plenty of full diffusion checkpoints, before you've even loaded the UNet. If you've read anything about the text encoder becoming "a second model you load and budget for" rather than an afterthought, this node is the sharpest example of it in the pack - the encoder stack here can outweigh the actual image model.

(One thing worth being clear on: this is HiDream's original text-encoder recipe - HiDream-I1/Full - not the newer pixel-space HiDream-O1, which drops CLIP and VAE loading entirely and doesn't use this node at all.)

Inputs and output

• clip_name1 through clip_name4 - your four encoder files from ComfyUI/models/clip, in the order your checkpoint expects (clip-l, clip-g, t5xxl, llama, per the recipe above).

Output is a single CLIP, same as any other CLIP loader - feed it into CLIPTextEncode.

That's the whole surface. No device picker, no offload slider on this particular node - just four filenames in, one CLIP out. Given how heavy this specific combination gets (that Llama file alone is the size of a mid-sized checkpoint), it's exactly the kind of loader you'd want offload controls on. If your install also has QuadrupleCLIPLoaderDisTorch2MultiGPU available, that's the sibling meant to carry device / virtual_vram_gb style controls for spreading this load out - check that one first if VRAM is tight.

Installing it

ComfyUI Manager: search "ComfyUI-MultiGPU," install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/pollockjj/ComfyUI-MultiGPU

No extra Python dependencies - it's a thin wrapper around ComfyUI's own loading code. Every node in the pack, including this one, has a built-in Help entry (the question-mark icon on right-click) with the pack's own documentation if you want it without leaving the app.

Common issues & troubleshooting

Out of VRAM before the sampler even runs. With four encoders - one of them an 8B LLM - this is the single most VRAM-hungry loader in the whole pack. If you're on a card under 24GB, expect to need offloading (CPU or a second GPU) for at least the Llama component; this plain loader doesn't give you that lever itself.

Wrong file in the wrong slot. All four inputs are just filename dropdowns with no type checking against what the model expects - loading them out of order, or substituting a plain CLIP-G export where "long clip-g" is expected, produces broken conditioning rather than an error. Match the exact files HiDream's model card specifies.

A file doesn't show up in the dropdown. Refresh the ComfyUI page after adding new files to models/clip - the node list doesn't pick up new files on disk without a reload.