Load CLIP (Triple)

Load CLIP (Triple) is the rare ComfyUI node whose entire audience fits in a small room: it loads three text encoders at once, and there's basically one reason to do that - Stable Diffusion 3. The node description says it in one line: sd3: clip-l, clip-g, t5. If you're not running SD3 or one of its close relatives, you almost certainly want Load CLIP (Dual) instead. This is the completionist's loader, and for SD3 it's the honest full setup.

How it works

Three required inputs:

• clip_name1 - the CLIP-L text encoder file.
• clip_name2 - the CLIP-G (OpenCLIP) encoder file.
• clip_name3 - the T5-XXL encoder file.

All three come from ComfyUI/models/text_encoders. The node loads them together and merges them into one CLIP object that produces a combined embedding - the same wrapper mechanism as the dual loader, just one encoder more. That single CLIP output feeds CLIP Text Encode (Prompt), and from there it's the standard road to conditioning and your sampler. Nothing else about the node differs from its siblings: no extra inputs, no tricks, no strength dials. Load three, encode, generate.

Why three encoders, and why SD3

SD3's architecture (the MMDiT that the whole field then copied) was designed to condition on all three encoder types at once. That's the "triple text encoder" story in the launch material, and it's genuinely how the architecture was supposed to run. The practical wrinkle the community found immediately: SD3 also accepts any two-encoder subset - clip-l + t5, clip-g + t5, clip-l + clip-g - through the Dual loader, at a meaningful VRAM saving. So the Triple loader is the "as designed" configuration, and Dual is the "still works, lighter" one. Which you choose is a VRAM-and-quality trade, not a correctness question.

Honest context, though: SD3 had a rough life. The June 2024 launch was famously broken, its ecosystem never formed, and by the time 3.5 arrived as a competent base, the community had largely moved on. So this node's practical footprint is small. That doesn't make it wrong - if a workflow you downloaded names it, it's load-bearing, and SD3.5 workflows do exist.

Gotchas

• Three files, one folder, easy mixup. All three encoders live in models/text_encoders, so it's very possible to load clip-g into the clip-l slot and get a CLIP that technically loads and semantically mumbles. Match the order to the recipe.
• T5 is the weight. T5-XXL dwarfs the other two. If you're short on VRAM, an fp8 T5 or a dual-encoder subset is the lever - the community's standard move, not an edge case.
• Don't reuse it for Flux. Flux wants exactly clip-l + t5; feeding it a third encoder through this node is a mismatch. Triple is for SD3, full stop.

It ships with ComfyUI core - no install. It's the node you load when the model card says "SD3.5" and you want the whole intended stack, and the node you skip when 12GB of VRAM says otherwise.