FluxDisableGuidance

Most diffusion models steer generation with a CFG scale knob. Flux doesn't - it bakes a guidance value into the conditioning itself as an extra embedding, and that value is normally set by CLIPTextEncodeFlux. FluxDisableGuidance is the kill switch: it removes that guidance embed completely, so the model gets no guidance signal at all.

Why you'd want that

The guidance embed exists because Flux was trained with it. But not every model in the Flux family wants it, and the distilled/step-checkpoint crowd in particular found that a stale guidance value actively hurt outputs - washed-out contrast, overcooked edges. When you run a distilled Flux (the 4-8 step finetunes that dominate local use), the guidance embed can fight the distilled sampler. This node strips it and lets the model run on raw conditioning.

It's also the aggressive sibling of FluxGuidance, which sets a custom guidance value (the 1000.0 default on some checkpoints is a classic cause of that "bleached" Flux look). If tuning the number doesn't fix it, removing the signal entirely is the next step.

How it works

One input, one output, zero knobs:

• conditioning in → CONDITIONING out.

The source is a single line: it sets guidance to None on every conditioning entry. The node doesn't touch your prompt embeddings, your timestep ranges, or anything else - it just erases the guidance field. Wire it between your text-encode node and the KSampler and you're done.

When not to bother

If your Flux checkpoint was trained with guidance (the original FLUX.1 dev and schnell lineage), removing it wholesale is usually a downgrade - that's why the node exists as an explicit, deliberate step rather than a default. The people who reach for it are running distilled or community-finetuned checkpoints where the guidance pathway is either meaningless or actively harmful. Related truth from the wider conditioning world: once you're at CFG 1 (which distilled models are), the negative prompt does nothing anyway, so the guidance embed is the only prompt-adjacent signal you have left to remove.

It ships with ComfyUI core - it's been there since the original Flux nodes landed in August 2024. If your Flux images look overcooked and fiddling with the guidance value in the text-encode node did nothing, this is the two-second experiment that tells you whether the guidance pathway is your problem. Add it, rerun, decide.