🐧 Sigmas Emphasis

A noise schedule is a list of sigma values, and normally it's generated once by a scheduler and never thought about again. Sigmas Emphasis is the node for the rare occasion you do want to think about it: it takes an existing SIGMAS schedule, resamples it to a new step count, and lets you cluster those new steps toward the beginning, middle, or end of the curve.

Why would you? Because where you spend your denoising steps is a real dial. More steps packed at the start of the schedule spends effort on structure and composition at high noise; more at the end spends it on fine detail and settling. On SD 1.5/SDXL-era models people reached for reshaped schedules (Karras) to do exactly this - concentrate work in the middle. On flow-matching models that trick mostly backfires, because the straight trajectory doesn't want aggressive reshaping, which is why Karras is now a known failure on those. This node is a more surgical version of the same idea: it keeps the original curve's geometry and just redistributes where the steps land, so it's the tinkerer's tool rather than a new default. If you're on a stock flow-matching model and happy with the output, you probably don't need it - but it's the right tool when you want a specific number of steps and a specific emphasis without changing sampler or scheduler.

How it works. The input sigmas are treated as a curve. The node places steps + 1 evenly spaced positions along it, then warps those positions with a cubic toward your chosen emphasis point, with emphasis_factor blending between even spacing (0.0) and full clustering (1.0). The first and last sigmas - including the trailing 0.0 that ends every schedule - are always preserved exactly, so you never lose the clean "zero noise" endpoint.

Inputs. sigmas (any schedule - straight from BasicScheduler or out of another scheduling node), steps (default 10; the output holds steps + 1 sigmas), emphasis (a dropdown: beginning, middle, or end), and emphasis_factor (0.0 to 1.0, default 0.0 - so out of the box it just resamples evenly, and the emphasis is opt-in).

Output. One SIGMAS socket, wired exactly where the original schedule went - usually into your sampler's sigmas input, or through more of this pack's scheduling nodes (Sigma Replace, Sigmas Smoother, and friends). It's a drop-in between the scheduler and the sampler.

A couple of grounded warnings, because this is the kind of node that invites fiddling. On flow-matching models the KB's advice is consistent: they want balanced sigma schedules and dislike aggressive redistribution - the big reshaped curves are what fall apart. So emphasis_factor near 1.0 is where you should be suspicious of artifacts, not where you should live. And remember the ancestral-sampler rule if you're testing: converging samplers (plain Euler, eta 0) are the only ones where changing steps actually refines instead of re-rolling the image.

Install. ComfyUI Manager → search "ComfyUI-SuperNodes", or git clone https://github.com/sonnybox/ComfyUI-SuperNodes into ComfyUI/custom_nodes and restart. No models, no dependencies beyond the pack's single matplotlib (which powers its sibling Sigmas Graph node - handy for actually seeing what your emphasis did to the curve).