Nodes/ComfyUI-Adept-Sampler/Adept Scheduler (JYS)
ComfyUI Node

Adept Scheduler (JYS)

The low-step schedule that jumps ahead on purpose

By nawka12·Created 8 months ago·Updated 4 months ago· 4
Adept Scheduler (JYS)
  • model
  • SIGMAS
steps20

"Jump Your Steps" is the answer this pack's author has to the classic low-step-count problem: at 8–15 steps you can't afford to spread your steps evenly, so you have to decide what to skip. JYS's strategy is explicit - take big jumps early, cluster steps in the detail-forming region, and save fine steps for the very end. It's a schedule designed for people who want decent output at step counts where other schedules visibly struggle.

The mechanism starts from timestep arithmetic rather than a sigma curve. The node computes a timestep sequence in the model's own 0–1000 space, structured in three blocks. The early phase takes aggressive jumps - from 1000 down toward 600 in big, chunky strides - because early denoising steps at very high noise are cheap to skip; the model mostly establishes broad structure there and a giant jump costs little. The middle phase, from 600 down to about 200, is where the schedule stops skipping: it computes two sub-blocks, one for structure (jumps from 600 toward 300) and one for detail (small jumps from 300 to 200), because this is the region where diffusion models make their most consequential decisions. The final phase walks fine steps from 200 down to 0 to polish the result. If the computed sequence comes up short of your requested step count, it fills gaps by bisecting between existing timesteps until it has exactly what you asked for. The whole thing then maps into sigma space via the Karras rho formula.

The practical upshot: compared to a uniform or Karras schedule at the same low step count, JYS spends a bigger share of its budget on the 300–200 detail zone, which is where visible quality lives at low steps. It's the pack's low-step specialist, and the README groups it with the "research-based" schedules alongside AYS-SDXL.

Inputs: model and steps (default 20), outputting SIGMAS into SamplerCustom. If you're running a distilled or Turbo-style checkpoint at 4–8 steps you're outside this schedule's design envelope (those want their own fixed-step schedules), but for normal checkpoints at 8–15 steps it's a strong candidate.

Install is the pack-wide one:

cd ComfyUI/custom_nodes
git clone https://github.com/nawka12/ComfyUI-Adept-Sampler

Restart, or install via ComfyUI Manager by searching "ComfyUI-Adept-Sampler". Pure Python, no requirements.txt, no model files.

The honest take: JYS is the classic "give it a fair shot before you judge it" node. Its name overpromises slightly - "Jump Your Steps" makes it sound like a free lunch, and it's not; at high step counts it converges toward behaving like a detail-biased schedule, and you won't notice a difference from the others. At low step counts, though, the jump logic does what it says, and that's exactly where most schedules fall apart.

Categorysampling/adept/schedulers

Inputs (2)

NameTypeDefaultDescription
modelMODEL
stepsINT201–10000

Outputs (1)

NameTypeDescription
SIGMASSIGMAS