CLIP Text Encode (HiDream)

HiDream-I1 is the 17B latent model that topped leaderboards when it opened under MIT in April 2025, and it's picky about how it takes a prompt: it wants four text encoders, not one. CLIPTextEncodeHiDream is the node with all four boxes. If you've only ever seen CLIP-L + CLIP-G + T5, the fourth box - a full LLM - is the part that feels new.

The four boxes

• clip_l - the small CLIP encoder. Short, keyword-style content.
• clip_g - the big CLIP encoder.
• t5xxl - the T5-XXL.
• llama - the LLM encoder. The loader's recipe names it: llama_8b_3.1_instruct.

HiDream-I1 was one of the first open image models to lean on an instruct-tuned LLM for prompt understanding, which is why it handles long natural-language instructions far better than the tag-driven generations of the SD era. The llama box is where the "a photo of X doing Y, shot on 35mm, with Z" sentence lives.

How it works

The node tokenizes each box into its own slot (l, g, t5xxl, llama) and returns a single CONDITIONING. It landed in core in April 2025 alongside the model, and it's paired with the QuadrupleCLIPLoader - the loader that combines the four encoder files into the one clip object this node expects.

The gotcha that bites everyone

The clip object has to contain all four encoders. If you load HiDream-I1's CLIP with a normal one-or-two-file loader and wire it in, the node errors or silently produces weak conditioning the moment it goes looking for a llama slot that isn't there. Use the Quadruple loader with the HiDream recipe, and you're set. The four text boxes also all default to empty, and an empty llama box is a bigger handicap than an empty T5 box on an older model - the LLM is carrying most of the semantic weight.

When to reach for it

If you're running HiDream-I1, this is the text-encode node - there's no simpler path. Fill all four boxes with the same prompt for the safe default, or treat them as separate lanes: keywords in clip_l/clip_g, the full instruction in t5xxl, the narrative in llama. A different story in each box is a legitimate lever, because each encoder contributes a different read of the same prompt to the final conditioning.

One thing worth knowing from the wider HiDream picture: the same company later shipped HiDream-O1 (the 8B pixel-space model), which has no text encoder at all and is prompted through a different node entirely (HiDreamO1ReferenceImages and friends). If your HiDream checkpoint isn't taking your prompt, check which model family you loaded - this node is for the I1 latent line, and it's not interchangeable.

Ships with ComfyUI core, no install.