LogSigmaStepSamplerNode

This is the twin of LogSigmaSamplerNode from the same pack, and the difference is one control: instead of choosing the active window as a percentage of the run, you choose it by step number. Same family, same core idea - feed the model a slightly adjusted sigma during part of sampling so it invents more detail or softens, the same technique Detail Daemon and the "Lying Sigmas" crowd use.

The mechanism is identical under the hood. It wraps your existing sampler and, for steps inside your window, passes the model a sigma that's linearly interpolated between the true sigma and the value from your LogSigmas curve, scaled by lerpfactor (default −0.05 - a 5% reduction). Outside the window, the model sees the true noise level.

The inputs that matter

• sampler - a SAMPLER to wrap (any sampler object from your graph).
• LogSigmas - your custom sigma curve, as SIGMAS. The reference curve it interpolates toward.
• lerpfactor - push strength, −0.999 to 0-ish. Negative reduces sigma (detail/sharpening), positive increases it (softening).
• start_step / end_step - the window, in absolute step indices. The source indexes into your sigmas tensor and applies the lie when start_step <= index <= end_step.

Output is a SAMPLER - drop it into a KSampler or sampler-custom node in place of your original.

Which one to use: percent or step?

Honestly, pick whichever your workflow thinks in. If your runs are a fixed step count and you think "I want this effect on steps 4 through 12," use this one. If your step count changes between runs and you think in fractions, the percent-based sibling keeps the window proportional. One practical trap: this node's window is an index into the provided sigmas, so if your sampler's step count doesn't match the length of LogSigmas, the window lands in a different place than you intended. Keep them aligned.

Install

Manager → search TBG Takeaways → install, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/Ltamann/ComfyUI-TBG-Takeaways

Under TBG/Takeaways after restart. No deps, no downloads.

Reasonable expectations: this is an experimental, one-dev Patreon node (Tobias Laarmann's "development takeaways"), so you're on your own for fine-tuning - the tooltips are the whole manual. Same gotcha as the sibling: some versions print debug lines to the console during sampling; harmless, ignore them. Start gentle (small lerpfactor), view at 100% zoom, and remember the window is what keeps the effect from wrecking your composition.