CLIPTextEncodeFlux

Flux is the model that taught a whole generation that one prompt box isn't enough. It reads your words with two text encoders at once - CLIP-L, the same small image-text aligner from the SD days, and T5-XXL, a much bigger language model that actually understands sentences - and this node is how you talk to both of them separately. The trick: you don't have to give them the same prompt. That's not a gimmick, it's the standard way people run Flux.

How it works

CLIPTextEncodeFlux tokenizes clip_l through the CLIP-L path and t5xxl through the T5-XXL path, then merges the two token streams into one conditioning. It also takes a guidance number and bakes it into the conditioning dict, which is how Flux's guidance-distilled architecture gets its "how hard to follow the prompt" signal without a traditional negative prompt. (No negative prompt here - Flux dev is guidance-distilled; if you want real CFG with negatives you need a de-distilled variant or the PerpNeg workaround.)

The inputs that matter

• clip_l - the CLIP-L prompt. The community convention is tag-soup here: comma-separated descriptors, style words, quality tags. It's the old-school half of the brain.
• t5xxl - the T5 prompt. Full natural language, whole sentences, spatial relationships. This is where Flux's famous prompt adherence comes from. Keep the sentence in one box and the tags in the other and you get both kinds of understanding.
• guidance - default 3.5, and that default is the single most disputed number in the Flux ecosystem. The "3.5 is too high" camp is large and loud: at 3.5, output can come out oversaturated or overcooked, and a lot of the "Flux looks bad/plasticky" complaints trace straight back to it. The standing community advice for the dev model: 1.8–2.5 with a well-written prompt, and remember higher resolutions tolerate lower guidance. With LoRAs stacked, people do push back up toward 4–4.5.

The single CONDITIONING output feeds your sampler's positive slot. There's no negative slot - that's by design, see above.

How you get it

Ships with ComfyUI core, listed under the Flux conditioning category. The heavy lifting is the models: the Flux checkpoint, plus both text encoders - the T5-XXL file (which is several GB and the main VRAM/disk tax) and the CLIP-L file - plus the Flux VAE. All in the usual models/ folders, and any of ComfyUI's official Flux example workflows shows the exact wiring.

Where people get burned

Wrong sampler/scheduler. The classic Flux complaint - muddy, blurry output - is almost always Euler + simple/normal not being used. That's not this node's fault, but it's the most common thing blamed on it. The 3.5 default. If your images look overdone or plastic, drop guidance before you rewrite the prompt. Treating both boxes the same. Paste the same giant sentence into both and you waste the CLIP-L channel; the whole point is that tags and prose pull in different directions. And one honest caveat: the dual-encoder trick is a Flux 1 era skill - the 2026 models moved to single LLM encoders, and this node only applies to the Flux 1 family.