MOSS DyPE-M Schedule (Info)

This node renders nothing. It doesn't touch a model, it has no latents, and if you wire it into a workflow expecting an effect on your image, you'll queue, wait, and get exactly what you had before. What it does is print you the schedule the patch node would use - which, at 8 MP and a handful of steps, is the difference between diagnosing a bad render in ten seconds and re-rolling seeds for an hour.

What it's for

MossDyPEPatch decides, per denoising step, how hard to extrapolate the position encoding. That strength is κ, it decays over the run, and you cannot see it. When an image comes out crunchy and over-detailed, or flat and under-resolved, you're left guessing whether the strength was too high early, too low late, or whether the schedule you picked does nothing at all in an 8-step run.

The Schedule node answers that as text. It builds a flow-matching sigma schedule from your step count and shift - the same curve ModelSamplingAuraFlow and friends produce - computes κ for each step, and evaluates the YaRN band frequencies so you can see band 0, band 8 and band 15 side by side. That last part is the bit nobody expects to be useful: watching the highest band move toward the base frequency is how you can tell the high-frequency restoration is actually happening rather than being asserted.

The inputs

steps and shift are the two that decide whether the preview resembles your run. steps should be your sampler's step count; shift should be your ModelSamplingAuraFlow shift (the pack's example workflow uses 5.0, which is the default here too). Get those wrong and the table describes a schedule you aren't running.

method, schedule, lambda_s, lambda_t and base_resolution are the same knobs as on the patch node, and you should set them to the same values. Then report hands you the table as a plain string on the node's single output.

Reading it without shooting yourself in the foot

Here's the trap nobody warns you about: this node's schedule defaults to hybrid, while the patch node defaults to adaptive. Leave both on defaults and your preview quietly describes a different schedule than the one that will run. Copy the value across. (aniso is a legacy alias for adaptive, kept in the dropdown so old workflows still validate - same behaviour, old name.)

The second trap is what the table can't see. It has no target_scale, no isotropic_scale, no start_percent, and no measured spectrum, so it can't tell you anything about strength on a non-square canvas or about the adaptive schedule's measurement pass. The band columns are evaluated at a fixed scale of 2.0, which is a reference point, not your canvas. Read the κ column for shape, read the band columns for direction, and don't read either as a prediction of what 3200×2560 will do.

And check where the string goes. The node declares itself an output node, so it runs when you queue - but it doesn't pop a preview widget and it doesn't print to the console. If you queue and stare at nothing, that's why: report needs to be wired into a text or display node in your setup. That's the same plumbing pattern as any informational node - cost of entry is one extra node, benefit is not debugging blind.

Install

Same pack, no separate install:

comfy node install comfyui-mossdype

Or clone into ComfyUI/custom_nodes/ and restart. Zero dependencies - pyproject.toml declares none, so it loads even on a machine with no MOSS-Image weights and no working sampler. Which is the point: it's a dry-run tool.

When to actually reach for it

Before your first 8 MP attempt with a new lambda_t, and any time you've dropped to a 4–8 step DMD workflow and something smells off. Few-step regimes are exactly where the paper's σ² schedule falls apart - sigma stays high until the last step, κ never decays, and the high-frequency restoration never fires. That's the bug the hybrid and adaptive schedules exist to fix, and this node is how you confirm the fix is happening on your step count and your shift instead of taking it on faith.

It's also the cheapest sanity check in the pack. Staring at numbers for ten seconds beats rendering an 8 MP canvas to discover your schedule was a flat line.