MD: Noise Decay Scheduler (Advanced)

Stock schedulers give you presets - Karras, linear, exponential - and they're fixed. This node lets you draw a noise decay envelope and hands it to the pack's PingPong samplers as a SCHEDULER object, plus a matplotlib plot so you can actually see what you made before you commit to a render. It's the same idea as an LFO or envelope in a synth: the shape you choose decides whether the process dives fast and smooths out, or lingers at the top then crashes down.

How it works

The core is a decay-curve generator. algorithm_type picks the math: polynomial (the standard ease-in, with decay_exponent controlling the curve - 1.0 is linear, 2.0 is quadratic), sigmoidal (an S-curve), piecewise (you type manual anchors like 1.0, 0.8, 0.2, 0.0 in custom_piecewise_points), fourier (an oscillating cosine wave, fourier_frequency sets the cycle count), exponential, and gaussian (a bell).

start_value and end_value pin the envelope's endpoints - 1.0 decaying to 0.0 is the standard shape, but a 0.0 start rising somewhere is your attack envelope. invert_curve flips the whole thing vertically (computes 1 - y), which is the fastest way to turn a decay schedule into an attack schedule. There's also enable_temporal_smoothing with a smoothing_window for a moving-average filter - genuinely useful in piecewise mode where the joins between your manual points can be jagged.

Two quality-of-life toggles: use_caching (stores identical curves in memory so repeated queue runs don't recompute) and preview_steps (plot resolution; 20–50 is plenty).

The outputs

scheduler is the one you care about - a SCHEDULER object you wire into the scheduler input of PingPongSamplerNodeFBG, Basic, or Lite. plot_image is the visualization, an IMAGE you can route to a preview node. And here's the critical thing to understand:

This is not a SIGMAS output. It won't plug into a stock ComfyUI KSampler.

It's a pack-specific object built for the pack's own sampler family. If you load a shared workflow and this node looks broken, that's usually the mismatch - it belongs in an MD PingPong pipeline, not in front of a vanilla sampler.

Installing it

Ships with MD Nodes; the plot needs matplotlib, which is in the pack's requirements:

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/MDMAchine/ComfyUI_MD_Nodes.git
cd ComfyUI_MD_Nodes
pip install -r requirements.txt

Or search MD_Nodes in ComfyUI Manager and restart.

Where it fits

Think of this as the "curve design" front-end for the PingPong ecosystem: the hybrid scheduler gives you the standard sigma math, GITS adapts step spacing, and this one is for when you want a bespoke envelope and a picture of it. The author's own guidance is sensible - polynomial at exponent 2.0 for standard diffusion ease-in, and don't reach for fourier unless you know you want oscillation, because non-standard shapes give unpredictable injection results. Start with the defaults, look at the plot, then bend the curve.