Adept Scheduler (AOS-V)

If you're running a v-prediction model - NoobAI vpred, or any SDXL fine-tune that switched to a velocity target for real blacks and dramatic lighting - this is the scheduler the pack points you at first. v-prediction models need their noise handled differently than the classic epsilon ones, and AOS-V ("Anime-Optimized Schedule for v-prediction") is a schedule tuned for exactly that family.

Mechanically it's a three-phase ramp, which is the thing that makes it different from a plain Karras curve. The first ~20% of steps map a steep 0.5-power ramp scaled to 0.6, so denoising starts aggressively on the structural phase. The middle 60% run a gentle, near-linear climb from 0.6 to 0.9 - that's where the fine structure forms and the schedule deliberately moves slowly. The final tail is a much steeper 3.0-power ramp up to 1.0, which concentrates the last steps on cleanup. Each phase exists because the author's model of v-prediction is that it spends its early steps establishing composition and its final steps resolving detail, and a schedule that treats those regions the same leaves one of them under-served.

Inputs are deliberately sparse: model (to read the checkpoint's own sigma_min/sigma_max) and steps (default 20). Output is a SIGMAS socket that feeds SamplerCustom's sigmas input. The README pairs it with Adept Solver (order 2, corrector on) for the full recommended v-prediction recipe, and that pairing is genuinely good - the converging solver plus a v-prediction-aware schedule gets you dark, dramatic output without the washed-out mid-tones that plague vpred models run on epsilon-era schedules.

Because it reads sigma bounds from the model itself, you can use it across SDXL and SD 1.5 fine-tunes without touching anything - though it's tuned with anime/vpred SDXL in mind. If you're not on a v-prediction model, AOS-ε is the sibling you want instead; using AOS-V on an epsilon model isn't broken, it's just not what the curve was shaped for.

Install is the pack-wide one:

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

Restart ComfyUI, or use ComfyUI Manager and search "ComfyUI-Adept-Sampler". No requirements.txt, no model downloads - pure Python.

The take: on a v-prediction SDXL model this is a schedule that demonstrably understands the model class it's for, which is rarer than it should be. If your vpred output has always felt a step behind, this is the first node from this pack to try. Just remember the v-prediction companion rule from the KB: keep CFG low-ish or add a rescale node - the schedule is only half the vpred story.