FluxGuidance

FluxGuidance is the node that makes Flux's weird CFG situation make sense. Unlike SD models, Flux doesn't use a classic negative prompt + CFG scale - it uses a guidance value that's baked into the conditioning as a "distilled guidance" number, and the standard sampler in ComfyUI reads it from there. This node is just the clean way to set that number. One input, one output, and it's the single most impactful setting you'll touch on a Flux workflow.

How it works

Mechanically it's a one-liner: it stamps a guidance value into the conditioning dict. The sampler then uses that value to scale the guidance embedding during denoising. The default is 3.5, which is the sweet spot the Flux team tuned for - the guidance embed is trained so 3.5 is "normal." Crank it toward 8-10 and Flux follows your prompt more aggressively; drop it toward 1 and you get more freedom and sloppier prompt adherence. There's even a sibling node, FluxDisableGuidance, for completely zeroing it (a common trick for distilled models that don't want guidance at all).

Inputs and output

• conditioning - the Flux conditioning from a Flux text encoder.
• guidance - float, default 3.5, range 0–100. Realistic working range is roughly 1–10; above that you're usually just fighting the model.

Output is the CONDITIONING with guidance set, straight into the sampler.

Where it fits

Ships with ComfyUI core, and it's been there basically as long as Flux has - it's one of the first nodes you meet on a Flux workflow, sitting between the text encoder and the KSampler. No model files involved; it touches only the conditioning. If your Flux outputs look generic and underprompted, this is the first knob people reach for after checking the prompt itself.

Common issues

The beginner trap is treating it like a CFG scale and pushing it to 8+, expecting more "obedience," then getting oversaturated, crunchy images with fried contrast. Flux guidance is not CFG - above ~4-5 it mostly adds contrast and stress, not prompt adherence. Distilled Flux models (Turbo, Schnell-style, and the quantized favorites) often want a lower value or none at all; leaving it at 3.5 can fight the distillation. And it does nothing on non-Flux models - the guidance key is Flux-specific, so this node is wasted on SD/SDXL pipelines. Also note: if your workflow was built before this node existed, it was probably hardcoding guidance via a ModelSampling node - this is the simpler, official replacement.