TT CLIP Text Encode FLUX

FLUX doesn't prompt like SDXL. It runs two text encoders - a small CLIP-L and a huge T5-XXL language model - and the T5 is why FLUX understands "a cozy cabin in the snow, cinematic light, shot on 35mm" as a sentence instead of a tag list. Stock ComfyUI's basic CLIP Text Encode hides that split. TT CLIP Text Encode FLUX gives you a separate box for each encoder, plus the guidance dial FLUX actually uses.

From the TenserTensor pack, this is the FLUX side of the SDXL encoder you may have seen from the same family. Same philosophy: expose the model's real controls instead of papering over them.

The inputs

• clip_l_positive and t5xxl_positive - two positive prompt boxes. CLIP-L is the keyword machine: "bokeh, rim light, 85mm, film grain." T5-XXL is the sentence machine: write the full scene in natural language. The pack's docs call the combination a feature, and the KB backs it up - that "short tags for CLIP-L, natural language for T5" split is how people actually get good FLUX results.
• clip_l_negative and t5xxl_negative - two negative boxes. Set them if you like, but read the next paragraph before you expect magic.
• guidance (default 3.5) - FLUX's guidance scale. This is the knob that matters, and the KB is clear: 3.5 gives you the "Fluxmaxxed" same-face look. For people, drop to 1.8–2.0; go higher only with a realism LoRA.

Outputs are POSITIVE and NEGATIVE conditioning, both wired for any sampler.

The honest FLUX reality

FLUX is guidance-distilled: its guidance is baked into the weights, and at low guidance the negative prompt box is largely inert - there's no unconditional pass for it to steer. So treat the negative boxes here as an insurance policy, not a control. The real levers are the positive prompts (write constraints as presence: "clean studio background, sharp focus") and the guidance dial. This isn't a flaw of the node - it's how the model works, and the node at least shows you the guidance control rather than hiding it.

Install and caveats

cd ComfyUI/custom_nodes
git clone https://github.com/tenser-tensor/ComfyUI-TenserTensor

or ComfyUI Manager → "TenserTensor" → install → restart. Deps: gguf, kornia.

The clip input needs a proper FLUX dual encoder (CLIP-L + T5-XXL) - load it with this pack's TT FLUX Models Loader or a DualCLIPLoader, or the encoder will error. And the standing pack caveat: this is a legacy V1 node, moved to Deprecated/ after the migration to ComfyUI's API V3, with removal planned for a future major release. It works fine today. If the guidance at 3.5 is giving you cloned faces and you'd forgotten the dial was there, that's the setting to touch first - not the sampler, not the seed.