☁️BizyAir CLIPTextEncodeFlux

What it is

Flux doesn't use one text encoder like most Stable Diffusion checkpoints - it uses two, CLIP-L and T5-XXL, at the same time, and it wants them fed separately rather than concatenated into a single prompt string. This node is BizyAir's cloud version of Flux's dedicated text-encode node, built specifically around that dual-encoder setup, as opposed to BizyAir_CLIPTextEncode (the plain single-text version elsewhere in this pack) which won't get Flux-shaped conditioning right.

You'd use this one specifically when your loaded model is a Flux checkpoint and you're running the rest of your chain through BizyAir's cloud sampling nodes.

How it works

CLIP-L, the smaller encoder, handles the overall gist and style; T5-XXL, the much larger language-model-style encoder, handles detail and comprehension of longer, more descriptive prompts. Flux was trained expecting both signals, so splitting your prompt across the two fields - rather than writing one string and letting a node internally duplicate it - is how the model was actually built to be conditioned. On top of that, Flux is a guidance-distilled model: the classifier-free-guidance correction that normally happens at sampling time (the CFG value in a KSampler) is baked into training instead, and applied here via the guidance parameter rather than a separate CFG scale downstream.

Inputs and outputs that matter

clip takes a BIZYAIR_CLIP - the output of a BizyAir CLIP loader, not a stock ComfyUI CLIP socket. clip_l and t5xxl are both multiline text fields - in practice, people often write the same or very similar prompt into both, though the encoders are free to diverge if you want to lean on T5's stronger comprehension for detail while keeping CLIP-L's field shorter. guidance (default 3.5, range 0–100) is Flux's built-in guidance strength - this is the knob that plays the role CFG normally plays elsewhere; 3.5 is Flux's own commonly-used default and a reasonable starting point before you push it higher for stronger prompt adherence or lower for more natural, less "AI-ish" output.

The output is conditioning, typed BIZYAIR_CONDITIONING - it feeds into a BizyAir guider or sampler node, not a stock ComfyUI KSampler.

Installing it

Install through ComfyUI Manager (search "BizyAir") or clone it yourself: cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git, then restart. A BizyAir API key is required before any node in this pack, including this one, does anything - the login prompt on first launch gets you set up.

Common issues

Using this node with a non-Flux model loaded upstream is the most common mismatch - Kolors, Wan, and other architectures don't have a CLIP-L/T5-XXL split and won't accept this conditioning correctly; pair it with BizyAir_CLIPTextEncode for non-Flux checkpoints instead. Setting guidance too high is the other trap people carry over from CFG habits on other architectures - Flux's guidance scale behaves differently, and pushing it well past the 3.5 default tends toward oversaturation and artifacts rather than better prompt following, the same failure mode distilled models generally show when you apply guidance corrections too aggressively.