Adept Scheduler (AkashicEQFlow)

Every scheduler in this pack's Akashic line is a different answer to the same question: what does an EQ-VAE SDXL model actually want from its noise schedule? AkashicAOS and its Alt warp a Karras ramp after the fact. AkashicEQFlow goes the other way - it constructs the step placement from scratch, around the single moment where the model does its most important work, and only then maps it into sigma space.

That moment is the structure-to-detail crossover, the point in the trajectory where log-SNR crosses zero and the model shifts from establishing composition to resolving detail. Diffusion models make their most consequential decisions there, and EQ-VAE's variant of that transition sits somewhere specific that a generic schedule won't know about. AkashicEQFlow builds a smooth probability density peaked around that crossover region, then uses CDF inversion to place steps - the peak of the density is where the steps bunch up. Because the density is continuous, the resulting step ratios come out smooth without needing the post-hoc clamping the other schedules rely on. Two practical guardrails sit on top: a ratio cap and slew-rate limiting, so a multi-step solver can't get handed a monstrous jump between adjacent steps.

The construction is also step-count-aware: the crossover center shifts slightly with the number of steps (more steps, the center drifts a bit toward the detail end), and the density keeps both tails alive so neither the early structure phase nor the final cleanup gets starved of steps. That's the "crossover-focused" claim in action - it's not that the schedule only cares about the middle, it's that the middle gets density proportional to how much the model cares about it.

Inputs are the bare pair: model and steps (default 20), outputting SIGMAS for SamplerCustom. It's the third option in the README's EQ-VAE recipe alongside AkashicAOS and AkashicAOS Alt, feeding AkashicSolver v2 (tau 0.5–0.6, order 2, external rescaleCFG 0.7). If you're already on EQ-VAE and the two AkashicAOS flavors both felt slightly off, this is the one with a genuinely different step distribution profile - the most distinct member of the trio.

Install is the pack-wide one, and it's as dependency-free as this pack gets:

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

Restart ComfyUI, or install via ComfyUI Manager by searching "ComfyUI-Adept-Sampler". No requirements.txt, no model downloads.

The honest take: like the rest of the Akashic line, this node is pointless unless your model actually uses an EQ-VAE - it's a niche schedule for a niche model family. But it's also the intellectually honest one: instead of warping an existing curve and hoping, it designs the step distribution around the mechanism that matters. For EQ-VAE users it's worth a serious A/B against the other two; for everyone else, the AYS or AOS schedulers are the ones to use.