Nodes/ComfyUI-MultiGPU/Load CLIP (Triple)
ComfyUI Node

Load CLIP (Triple)

SD3's three-encoder loader

By pollockjj·Created 2 years ago·Updated 3 months ago· 925
Load CLIP (Triple)
    • CLIP
    clip_name1
    clip_name2
    clip_name3

    SD3 is the one architecture in common use that wants three separate text encoders loaded together - clip-l, clip-g, and T5-XXL, per the node's own description: "sd3: clip-l, clip-g, t5." TripleCLIPLoaderDisTorch2MultiGPU is built for exactly that: three file pickers - clip_name1, clip_name2, clip_name3 - combined into one CLIP output.

    Why three at once instead of stacking two like Flux does: SD3 kept both of SDXL's CLIP encoders (clip-l and clip-g) and added T5-XXL on top rather than replacing them, on the theory that the extra encoder adds language understanding without giving up what the smaller CLIP models are good at. That's more text-encoding horsepower than almost anything else you'll load in ComfyUI, which is exactly why a dedicated triple-loader - rather than three single loaders wired together by hand - earns a place in this pack.

    A note on what's actually on this node's page right now

    Worth being upfront about: unlike its siblings elsewhere in the DisTorch2 family, the inputs currently exposed on this node are just the three CLIP pickers - no compute_device, virtual_vram_gb, or donor_device controls show up alongside them, the way they do on every other DisTorch2 loader in this pack. If you came here expecting the same offload sliders the UNet and single-CLIP DisTorch2 nodes have, you won't find them on this one today. That's a real inconsistency with the rest of the family, which all expose that pattern - so if splitting this specific loader across devices matters to your workflow, it's worth checking you're on the latest release of the pack before assuming the controls are gone for good.

    What it does regardless

    It loads SD3's three encoders together in a single node instead of three separate loaders, producing one properly-combined CLIP object ready for SD3's text-encode step.

    • clip_name1, clip_name2, clip_name3 - by SD3 convention, clip-l, clip-g, and t5xxl respectively.
    • Output: CLIP → your text-encode node.

    How to install it

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

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

    then restart. No extra dependencies.

    Common issues & troubleshooting

    This is a niche node in traffic terms. SD3 itself is a smaller slice of what people run today than Flux or Wan, so you'll mostly need this if you're specifically working with SD3-family checkpoints.

    Slot order. Match clip_name1/clip_name2/clip_name3 to clip-l/clip-g/t5xxl respectively - get it backwards and the node loads without erroring, but conditions your prompts wrong, which is a much harder bug to notice than a load failure.

    Need device control and don't see it here? The plain single-file loaders - CLIPLoaderMultiGPU and CLIPLoaderDisTorch2MultiGPU - do expose full device and offload control. You can load clip-l, clip-g, and t5xxl as three separate CLIPLoaderDisTorch2MultiGPU nodes and combine them with a CLIP-merge node if your workflow needs per-file placement this triple-loader doesn't currently offer.

    T5-XXL is the heavyweight of the three. clip-l and clip-g are both comparatively small; T5-XXL is the one that actually pressures your VRAM budget. If you're VRAM-constrained specifically because of this loader, a quantized (fp8 or GGUF) T5-XXL file is usually a faster fix than restructuring your whole graph around per-file device placement.

    Categorymodel/loaders

    Inputs (3)

    NameTypeDefaultDescription
    clip_name1COMBO0 options:
    clip_name2COMBO0 options:
    clip_name3COMBO0 options:

    Outputs (1)

    NameTypeDescription
    CLIPCLIP