Load CLIP (Triple)

If you're running Stable Diffusion 3, you've hit the fact that it doesn't use one text encoder or even two - it uses three: CLIP-L, CLIP-G, and T5. TripleCLIPLoaderMultiGPU is ComfyUI-MultiGPU's registration of that exact loader, so it can sit alongside the pack's other MultiGPU nodes and be picked from the same "multigpu" category instead of you hunting for the stock version buried in the regular loaders list.

Why SD3 needs three encoders

SD3 was the model that made the text-encoder stack itself a design decision rather than an afterthought. Two CLIP encoders inherited from the SDXL era, plus a T5-XXL borrowed from the language-model world, feeding three separate conditioning streams into the same diffusion transformer. It's the direct ancestor of what Flux later did with just CLIP-L + T5, and worth knowing: SD3's own paper found that dropping T5 entirely at inference barely dented general image quality - the real cost was in typography, rendering text in the image. That's why some lighter SD3 workflows load only clip-l and clip-g and skip T5 (using DualCLIPLoader's sd3 recipe) when they don't need legible text and want to save the VRAM. This node, TripleCLIPLoaderMultiGPU, is for when you want the full three-encoder setup.

Inputs and output

Straightforward, three required fields, no defaults to fuss over:

• clip_name1, clip_name2, clip_name3 - your three encoder files, pulled from ComfyUI/models/clip. Order them the way your SD3 checkpoint expects (typically clip-l, clip-g, t5xxl).

Output is a single CLIP - wire it into your CLIPTextEncode node the same way you would with any other CLIP loader.

One thing to know going in: unlike its DisTorch2 siblings in this pack (the Dual CLIP and VAE loaders), this particular node doesn't currently surface a device selector or the virtual_vram_gb offload controls - just the three filenames in, one CLIP out. If you need to steer where SD3's text encoders sit in memory, look at whether your ComfyUI-MultiGPU install has picked up TripleCLIPLoaderDisTorch2MultiGPU as well (the pack lists both) - that's the variant meant to carry those controls.

Installing it

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

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

No extra dependencies - this pack wraps ComfyUI's own loader nodes rather than shipping a separate inference library. Once it restarts, the node appears under "multigpu" in the node menu, and every MultiGPU/DisTorch2 node ships its own built-in Help entry (right-click the node → the question-mark icon) if you want the pack's own writeup without leaving ComfyUI.

Common issues & troubleshooting

Only three encoders and SD3 won't load / weird conditioning. Double-check the files actually match the checkpoint's expected trio - a common mistake is loading a CLIP-G file that's actually an SDXL export rather than the one paired with your specific SD3 checkpoint. If the checkpoint page lists specific encoder files, use those exact ones.

You expected a device dropdown and don't see one. As above - this is the base MultiGPU registration of the triple loader, not the DisTorch2 offload variant. If VRAM is the problem, either the pack's DisTorch2 triple-CLIP node (once available in your install) or trimming down to DualCLIPLoader's two-encoder SD3 recipe (clip-l + t5, or clip-l + clip-g) are the ways to actually reduce what's loaded.

Files not showing up in the dropdowns. Refresh the ComfyUI browser tab after adding new files to models/clip - new files on disk don't appear until the page reloads its node definitions.