🔧 Guidance Timestep (experimental)

Flux doesn't use classifier-free guidance the way SD1.5 and SDXL do. It's a distilled model - the teacher's CFG behavior got baked into the student during training, which is exactly why running normal CFG on top of a distilled model tends to oversaturate and over-contrast: you're applying the correction twice. Instead, Flux takes a single "distilled guidance" value through a guidance-embedding, usually set once and left flat across the entire sampling run. Guidance Timestep turns that flat value into a schedule - applying it only within a window of the sampling process instead of the whole way through.

How it works

value (0–100, default 2 - right in Flux's normal guidance range) sets the guidance strength. start_at and end_at (0–1, defaulting to 0.2 and 0.8) are fractions of the total sampling trajectory, not absolute step counts - the guidance value only applies between those two points. The idea: skip guidance early while broad composition is still forming, apply it through the middle where prompt adherence actually gets locked in, then ease off again late, where guidance mostly just adds contrast and saturation rather than helping. It's the same underlying idea as scheduling CFG on SD1.5 workflows, aimed at Flux's distilled-guidance mechanism specifically.

Inputs and outputs

model, value, start_at, end_at. Output: MODEL, patched - it goes into the same slot in your graph as any other guidance or model-patch node, right before the sampler.

Installing it

Via ComfyUI Manager, search "ComfyUI Essentials". Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cubiq/ComfyUI_essentials

Restart. This one is marked "(experimental)" in its own name, and the pack overall has been maintenance-only since April 2025 - don't expect this to graduate out of experimental status or get iterated on further.

Common issues & troubleshooting

Output looks the same as not using this node at all. Double-check that start_at/end_at actually cover the part of the sampling process you think they do - these are fractions of steps, not absolute step counts, so a 10-step run and a 40-step run apply this node at very different absolute step numbers even with identical start_at/end_at values. On a very short run, a narrow window like 0.2–0.8 might barely register.

Not sure what window is "correct." There isn't a widely-agreed schedule for this the way there is for CFG scheduling on older architectures - it's marked experimental for a reason. Treat any specific start/end you pick as a hypothesis to test on your own images against a plain flat-guidance baseline, not a settled recipe.

Results look worse than just using a normal flat guidance node. That's a real possible outcome, not necessarily user error - this replaces a well-understood, well-tested default (flat guidance) with an unproven schedule. If it's not helping on your workflow, going back to flat guidance is a perfectly reasonable conclusion.