Nodes/ComfyUI-SuperNodes/🐧 Sigmas Smoother
ComfyUI Node

🐧 Sigmas Smoother

Put the final denoise step back on a ramp instead of a cliff

By sonnyboxΒ·Created 11 months agoΒ·Updated 4 days agoΒ· 12
🐧 Sigmas Smoother
  • sigmas
  • SIGMAS
β—„smooth_steps1β–Ί
β—„interpolation_typelinearβ–Ί

Every sigma schedule ends with a drop to zero - the final step where the sampler goes from "still a bit of noise" to "clean image." On a coarse schedule that drop can be a cliff: one step is 0.3, the next is 0.0, and the sampler has to bridge a real gap in a single pass. This node turns that cliff into a ramp by inserting intermediate steps before the final zero, and the pack's README frames it exactly that way: "add an intermediary step(s) before full denoise." It's the most practical of the 🐧 SuperNodes scheduling family for people who aren't running a 30-step safety net.

How it works

  • sigmas - the schedule, typically ending in 0.0.
  • smooth_steps - how many extra steps to insert between the last non-zero sigma and zero (1–100).
  • interpolation_type - linear spaces them evenly from the last non-zero sigma down to zero; decay scales them down the aggressive way (last value Γ·2, then Γ·3, then Γ·4…), packing most of the extra work right at the end. Think of linear as "spread it out" and decay as "nudge, nudge, then done."

The node keeps your original steps untouched, drops the terminal zero, appends the interpolated values, and re-appends 0.0 so the sampler still gets a proper clean end state. Output is a longer SIGMAS schedule - more steps, same starting noise, same ending noise.

When you'd reach for it

  • Low-step schedules (8–12 steps) where the final jump to clean is visibly degrading fine detail. One or two smooth steps at the end is the classic cheap fix.
  • Distilled and fast models that use few steps, where the tail of the schedule does most of the damage. Note the community's general guidance - on flow-matching models, gentle schedule edits like this are the kind that help rather than hurt, because they don't redistribute effort across the whole curve.
  • Refining the last stretch of a custom schedule before you run the expensive render - pair with the pack's Sigmas Graph to see what you're about to do.

Install

ComfyUI Manager, search "ComfyUI-SuperNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/sonnybox/ComfyUI-SuperNodes

Restart ComfyUI. No model downloads. The pack's only extra dependency is matplotlib (used by Sigmas Graph). This pack targets ComfyUI's newer extension API - update ComfyUI if the nodes don't load.

Gotchas

  • Each smooth step is a real denoising pass: smooth_steps: 10 on a 20-step schedule is a 50% slower render. One to three is the sane range; beyond that you're paying sampler time for marginal refinement.
  • decay isn't "less aggressive," it's the opposite in the tail: it crams extra steps into the very end. If you wanted even spacing, that's linear.
  • It always appends 0.0 regardless of whether the input ended in it. Feeding a schedule that doesn't end in zero gets one added - generally the right call, but know it's happening.

It's a one-idea node, but it's a good idea. If your low-step outputs have that slightly unfinished "last step was too big" look, this is the cheapest fix in the pack.

CategorySuperNodes/Scheduling

Inputs (3)

NameTypeDefaultDescription
sigmasSIGMASThe input sigmas tensor, typically ending in 0.0.
smooth_stepsINT11–100Number of additional smoothed steps to insert between the last non-zero sigma and 0.0.
interpolation_typeCOMBOlinearMethod to calculate the intermediate sigma values.

Outputs (1)

NameTypeDescription
SIGMASSIGMASβ€”