CFG and guidance

How strictly the model follows the prompt, and the separate "guidance" number of Flux.

#In plain words

CFG (classifier-free guidance) is the cfg field of a sampler. At each step the model predicts twice: once with your prompt, once with the negative prompt (or with nothing). The final step moves from the second prediction towards the first, multiplied by cfg. A larger number means the picture follows the prompt more strictly.

• Too low: a limp, washed-out picture that follows the prompt loosely.
• Too high: burnt colours, harsh outlines, artifacts.
• Exactly 1: the second prediction is skipped. The step is twice as fast, and the negative prompt does nothing.

Guidance in Flux dev is a different number. Flux dev was distilled: it learned the effect of CFG and takes the desired strength as an input, written into the conditioning by the FluxGuidance node (guidance, around 3.5). The sampler's own cfg stays at 1. People call this "distilled CFG".

#Where you meet it in a graph

• cfg on KSampler, KSamplerAdvanced, SamplerCustom, CFGGuider, and the samplers of packs.
• guidance on FluxGuidance (Flux.1 dev and models built on it).
• Two-pass video graphs have a cfg on each sampler.

#Values people use

The model's page is the final word: authors give the values they tested.

#Typical mistakes

• A burnt picture: cfg too high for the model, or a speed-up LoRA with cfg above 1.
• Raising both FluxGuidance and cfg on Flux dev: harsh pictures and double time.
• "The negative prompt does nothing": cfg is 1.

#How the app helps

The lessons check cfg ranges (1 to 4 for Wan's two passes, exactly 1 with a speed-up LoRA) and say why. Check says when cfg 1 leaves a negative prompt without effect. In Properties, a field with a built-in note has a button What this field does next to its name; it opens the note under the field: what the field does and what to put in it.

#See also

• The negative prompt
• Sampler and sampler names
• Steps
• Speed-up LoRAs
• Flux.1