Adept Scheduler (AkashicAOS)

Here's a schedule that exists because a specific niche got specific. AkashicAOS is built for EQ-VAE SDXL models - checkpoints that use the EQ-SDXL-VAE family (from KohakuBlueleaf and the Feb 2025 EQ-VAE paper) rather than the stock VAE. Those VAEs trade a bit of compatibility for a meaningful jump in detail and color fidelity, but they change the latent statistics, and the old schedules don't respect that. This node is one of very few schedules tuned for the difference.

The mechanism is a "detail-progressive" curve. Where a plain Karras schedule distributes steps with a fixed power, AkashicAOS applies a progressive warping - a 0.85-power detail bias that shifts density toward the low-sigma detail phase - and then layers a sinusoidal mid-boost (strength 0.08) around the middle of the trajectory to reinforce the structure-to-detail transition region where models make their most consequential decisions. The whole ramp is normalized back to [0,1] and then mapped through the standard rho=7 Karras formula using the checkpoint's own sigma bounds. Step ratios get smoothed and capped so a multi-step solver doesn't trip over a giant jump between adjacent steps.

Inputs: model and steps (default 20). That's it. Output is SIGMAS for SamplerCustom's sigmas input. The README's EQ-VAE recipe is AkashicAOS (or its Alt/EQFlow siblings) feeding AkashicSolver v2 at tau 0.5–0.6, order 2, plus an external rescaleCFG node at 0.7. The rescaleCFG part is the part people skip and then blame the schedule for.

The honest framing: if you're not on an EQ-VAE model, this node does nothing for you. It's not "better" - it's specific. On a normal SDXL checkpoint, AkashicAOS is a mildly unusual curve that mostly behaves like a detail-biased Karras; you won't be harmed, but you won't see the magic either. It earns its keep only when your model is actually using an EQ-VAE, where the README's "detail-progressive" claim shows up as crisper fine structure and better colors through the full schedule.

Install is the pack-wide one:

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

Restart ComfyUI, or grab it via ComfyUI Manager by searching "ComfyUI-Adept-Sampler". No dependencies beyond torch, no model downloads.

One compatibility note: EQ-VAEs aren't drop-in with old diffusion checkpoints - the paper's own framing warns they're incompatible with models trained against the stock VAE. So this schedule assumes the whole model stack was built EQ-VAE-aware. If your EQ-VAE workflow has been fighting for good results, start here rather than the general-purpose schedules; it's the one that knows what your VAE is.