MF Graph Plotter

Every serious parameter sweep eventually hits the same question: what actually changed as I moved the slider? MF Graph Plotter (MF_GraphPlotter) is the node that answers it without you exporting to a spreadsheet - it plots a running line graph on the canvas, one point per execution, and keeps the history even across ComfyUI restarts.

It's from MF PipoNodes (pierreb-mf / Moment Factory), and it's the most frontend-heavy node in the pack. The Python side is tiny: it records one integer per run. Everything you see - the chart, the reset button, the save button - lives in the JavaScript frontend.

How it works

Here's the part where the README and the shipped code disagree, and the code wins: there is no X input. The README talks about feeding both X and Y coordinates, but the actual INPUT_TYPES define only Y (an INT). The frontend auto-increments X for you - X is simply the count of data points plus one. You feed it one number per execution and get a time series. That's it.

The mechanism:

• Each execution appends your Y value to a per-node series stored in graph_plotter_state.json inside the pack's folder. That's real persistence - restart ComfyUI and the curve is still there.
• The chart itself is Chart.js, loaded from a CDN on startup. No bundled library, so the first load needs internet.
• The Reset Graph button posts to the pack's /graph_plotter/reset API endpoint and clears that node's data only (state is keyed by node ID, so two plotters stay independent).
• Save Graph exports the canvas as a JPEG at 0.95 quality. The README says PNG; the code says JPEG. Trust the code.
• The node is an output node with no STRING/INT outputs - you can't wire the chart further, which is fine; it's a visual instrument, not a data pass-through.

What it's actually good for

The pack's own examples are the right ones: CFG value on one axis and a quality score on the other to find a sweet spot, or loss/step for training runs. In practice, keep it simple - track one scalar per workflow, like a CFG sweep where Y is your manual quality rating:

CFG scale → KSampler ──▶ Save Image
    └────────────▶ MF Graph Plotter (Y)

Queue ten configs, rate each, and the shape of the curve tells you where quality peaked.

Install

cd ComfyUI/custom_nodes/
git clone https://github.com/pierreb-mf/ComfyUI-MF-PipoNodes
cd ComfyUI-MF-PipoNodes
pip install -r requirements.txt

Restart, or use ComfyUI Manager → "MF PipoNodes". Only aiohttp and pyyaml are required - no models, and Chart.js comes from the CDN at runtime.

Gotchas

• No internet on first launch = no chart. Watch the browser console for Failed to load Chart.js. Offline ComfyUI boxes get a graph that never draws.
• The node always executes (its change-detection returns NaN), so every queue adds a point - accidentally re-queueing a finished workflow appends junk points you then have to reset.
• Reset is per-node and via the API. If the pack's server routes failed to register (you'd see a warning in the console at startup), the Reset button does nothing.
• State file lives in the pack folder - if you wipe custom_nodes/ComfyUI-MF-PipoNodes to reinstall, your plotted history goes with it.