LC Sigma Resample

Most ComfyUI users never think of their noise schedule as something you can edit. You set a step count, you pick a scheduler name, done. But the steps are never spent evenly: the early high-sigma steps rough in composition and the late low-sigma ones add detail, and when a run needs more budget it's almost always one of those ends that wants it, not the middle. LC Sigma Resample is the node for exactly that moment.

It's from lonecatone23's LC123 pack, which bundles a whole family of sigma/split-sampling tools. This one is the smallest of them: SIGMAS in, SIGMAS out, two number widgets. You hand it an existing sigma curve - the list of noise levels a sampler steps down through - and it re-grids that curve onto a new step count while leaving both endpoints exactly where they are. Same σ path, new real step count, as the author puts it. Feed a 30-point schedule a multiplier of 2 and you get a 60-point version of the same path, and the sampler downstream genuinely runs those 60 steps. This is not cosmetic; it changes what actually executes.

How it works

The math is two parts. First it computes a target count:

new_steps = round(old_steps * multiplier) + adder

where old_steps is how many steps the incoming curve represents (points minus one). Then it linearly interpolates the old curve to land new_steps + 1 points along it, so the shape of the path is preserved and the start and end sigma don't drift. That's it - no model involved, no resampling of anything else. If the node ever looks like it's doing nothing, check that you actually changed multiplier or adder; defaults are 1 and 0, which is a pass-through by design.

The inputs that matter

The schema is refreshingly short:

• sigmas (SIGMAS, required) - whatever schedule you want denser or sparser.
• multiplier (default 1.0, range 0.05–20) - the tooltip says it plainly: "multiply the current step count first." 2.0 doubles the steps on this band; 0.5 halves them.
• adder (default 0, range −100…200) - then add or subtract a fixed number of steps, handy when you want +4 rather than exactly double.

Output is a single sigmas (SIGMAS) socket. Wire it into the sigmas input of an Advanced sampler - a KSampler (Advanced) or SamplerCustomAdvanced - the way the pack's own dual-sigma example workflows do.

Where it goes (and where it doesn't)

The pack's rule of thumb is worth memorizing: put it after a split, on the slice you want denser, not in front of the splitter. In a split-sigma setup you've divided one schedule into a high band (coarse, early) and a low band (detail, late), often so two different models each own part of the job. If you resample before the splitter you're densifying the whole schedule, which defeats the point - you might as well just raise total steps. Resample one of the split outputs instead and you change the density of that band only: extra steps for your detail model without making the coarse pass run longer. That's the "same as changing density on that band only" behavior the README describes, and it's a genuinely nice trick when one of your two models needs twice the budget of the other.

Installing and troubleshooting

The pack installs like any other: ComfyUI Manager → search "ComfyUI_LC123_nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes

then restart ComfyUI. For these sigma nodes there are no extra pip packages and no model downloads - the pack runs on ComfyUI's own env (torch/numpy). If it won't load, make sure __init__.py sits directly in custom_nodes/ComfyUI_LC123_nodes/ rather than nested one folder deeper, which happens when you unzip over an existing clone.

Two gotchas to dodge. First, a normal KSampler (the one with the steps widget and scheduler dropdown) computes its own sigmas internally - this node only matters where sigmas flow as a wire to an Advanced sampler. Second, remember the sampler really runs the new count. That's the feature, but it's also why a careless multiplier of 4 can turn a 20-step workflow into an 80-step one; halve a band, don't explode it, and you'll save real time where the extra steps earn their keep.