☁️BizyAir FluxGuidance

Flux doesn't use classifier-free guidance the way SD 1.5 and SDXL do. It's guidance-distilled - the model was trained with an approximation of CFG's extrapolation baked directly into the weights, so there's no separate unconditional pass to run. This node is where that baked-in guidance strength actually gets set: not as a sampler setting, but as a value attached to your conditioning before it ever reaches the sampler.

How it works

Normal CFG runs the model twice per step - once with your prompt, once without - and extrapolates away from the unconditional result to sharpen adherence. That doubles render time, and Flux was never trained to need it. Applying real CFG on top of a Flux graph anyway just burns time for nothing, which is exactly the mix-up the community complains about most: "You wrote CFG when you should have written Flux guidance... They are two different things and CFG doubles the time for generation." This node's guidance value is the actual lever - it patches the conditioning directly, and the sampler runs a single pass regardless of what you set it to.

3.5 is the widely-used default, though plenty of people run lower - 1.8-2.5 - for less of the oversaturated, "plastic" look Flux Dev tends toward at default settings, especially without a LoRA in the mix. With realism LoRAs loaded, people report pushing it back up to 4-4.5 without the same downside. Past roughly 6-8 you're mostly trading naturalism for harsher, more artifacted prompt-following.

The inputs and outputs that matter

• conditioning - a BIZYAIR_CONDITIONING, required. This has to come from a BizyAir text-encoding node upstream, not a plain CONDITIONING from a non-BizyAir CLIPTextEncode.
• guidance (default 3.5, range 0-100, step 0.1) - the whole point of the node.

Output: BIZYAIR_CONDITIONING, patched with the guidance value, ready for your BizyAir sampler.

How to install it

ComfyUI Manager: search BizyAir, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir

then restart. BizyAir wants an account and API key set up on first use per the README.

Common issues & troubleshooting

Node won't connect to your conditioning. It expects BIZYAIR_CONDITIONING specifically - make sure your prompt is encoded through a BizyAir text-encode node upstream in the same cloud graph, not a standard local CLIPTextEncode.

Trying to "boost" prompt strength by also raising sampler CFG. Don't. On a guidance-distilled model like Flux, real CFG on top of this node's value doesn't stack usefully - it doubles render time and tends to produce worse results, not stronger ones. Raise guidance here, leave the sampler's CFG alone.

Skin looks plastic/oversaturated. Classic symptom of running the 3.5 default without a realism LoRA - try dropping toward 1.8-2.5 and see if it clears up before reaching for anything else.