Adept Scheduler (AYS-SDXL)

"Align Your Steps" is NVIDIA's CVPR 2024 finding that the standard way people chose step counts for SDXL was wrong - instead of just subdividing a uniform schedule, you should place steps at specific, optimized fractions of the trajectory, and those fractions depend on how many steps you're running. This node brings that to ComfyUI with the actual paper's numbers baked in. It's the most defensible default in this pack for plain SDXL work.

The mechanism is refreshingly un-mysterious. For the common step counts - 10, 15, 20, 25, and 30 - the node uses exact preset sigma tables straight from the AYS work: a 10-step run places steps at 0.875, 0.750, 0.625… of the way down rather than evenly spacing them, because that's where SDXL actually needs them. For any other step count it interpolates the same idea onto the checkpoint's sigma range. Notice the presets are slightly detail-weighted - steps bunch toward the end of the schedule compared to a uniform ramp, which is the whole AYS thesis: spend your last steps where the image is being finalized.

Inputs: model and steps (default 20). Output: SIGMAS into SamplerCustom. The README recommends AYS-SDXL as one of the two scheduler choices for v-prediction SDXL models (alongside AOS-V), paired with Adept Solver at order 2 with the corrector on. It also works perfectly well on epsilon SDXL - it's the pack's "generic but principled" option.

The practical angle: if you've been on plain Karras and have felt like 20 steps is never quite enough, AYS often squeezes more out of the same count because its step placement is deliberate rather than geometric. It's also the schedule people reach for when a new SDXL checkpoint drops and they want a known-good baseline before they start experimenting.

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". No requirements.txt, no models, pure Python.

The honest take: AYS is exactly as good as the paper says, which is to say "a solid improvement over naive uniform spacing at low step counts, marginal at high ones." At 20+ steps, Karras and AYS are trading punches and you won't reliably tell them apart. Where this node genuinely stands out is low-step work - 10 to 15 steps on SDXL - where the preset tables earn their keep. For that, it's the safest pick in the pack, and it's the one I'd hand someone who just wants a good schedule and no drama.