analysis-graphing

name	analysis-graphing
description	RenoDX workflow for creating readable analysis graphs and plots from shader math, CSVs, EXRs, LUTs, hue sweeps, tone curves, gamut comparisons, energy/scalar maps, and test-pattern statistics. Use when graphing, plotting, visualizing, comparing curves, making dark-theme matplotlib figures, or avoiding repeated one-off plot scripts.
argument-hint	data source, variables to compare, output path, units/axis scale, and whether the graph is scratch or durable

RenoDX Analysis Graphing

Use this small skill for plots and graphs. Larger skills should call this out instead of embedding graphing rules.

Boundaries

• Focus on visualizing data, not generating source test images or editing shaders.
• Keep one-off plots and source scripts in a scratch output path unless the graph becomes a durable analysis artifact.
• Promote repeated graph workflows into tools/analysis/ with the data export beside the image.
• Prefer matplotlib for Python analysis unless an existing script in the same workflow already uses another plotting library.
• Use bt2020-png-generation for final HDR PQ PNG writing and signaling.
• Use hdr-test-pattern-generation for ramps, sweeps, charts, and synthetic image inputs.

Default style

• Use a dark theme by default: plt.style.use("dark_background").
• Save readable static images: usually PNG at dpi=150 to 180.
• Choose figure sizes for labels first, not minimum pixels; common overview plots are 14x9, 16x11, or 17x11 inches.
• Close figures after saving to avoid leaking state in batch scripts.
• If the graph will be inspected in an issue or PR, prefer a single self-contained overview image plus any focused split images.

Plot checklist

For every generated graph, make the output self-describing:

• Title states the experiment and transform/version being compared.
• Axes include units: nits, linear RGB, PQ code value, hue degrees, stops, frame index, etc.
• Legends use stable method names matching CSV column names or shader function names.
• Include reference lines for anchors such as zero, diffuse white, mid-gray, 1.0, peak nits, or gamut boundary when relevant.
• Do not normalize silently. If data is normalized, show the normalization factor in the title, label, or CSV.
• Use log/stops axes only when the labels make the scale obvious.
• Save the plotted source data as CSV when values are generated rather than loaded from an existing CSV.

Common RenoDX graph types

Graph type	Use for	Notes
Tone/inverse diagnostic curve	Vanilla vs RenoDRT/PsychoV/ACES matching	Mark diffuse white, mid-gray, peak, and shoulder anchors; inverse plots are for fitting/diagnosis, not final-frame inverse-tonemap strategy.
Gain/loss or delta plot	Comparing old/new math or fitted curves	Plot absolute output and error/delta, not only one.
Hue sweep	Gamut compression, hue preservation, negative-channel checks	Hue degrees on x-axis; include min/max channel or perceptual metric.
Gamut scatter / chip grid	BT.709, BT.2020, AP1/AP0 comparisons	State source gamut and adaptation path.
LUT/stat overview	LUT pair comparisons, channel summaries, error histograms	Keep CSV summaries beside the graph.
Image diagnostic panel	EXR/test-pattern before/after comparisons	Use fixed scales when comparing panels.

Existing examples

• Dark plot template for a minimal reusable matplotlib setup.
• tools/analysis/plot_cp2077_* for readable dark-theme multi-panel graphs.
• tools/analysis/validate_mb_compress.py for hue sweep validation and CSV-plus-plot output.
• scratch zelda_* curve and derivative plots.

Common mistakes to avoid

• Do not output light-theme graphs unless the user explicitly requests that style.
• Do not crop legends, tick labels, or colorbars; use tight_layout() or explicit layout spacing.
• Do not mix scene-linear, display-linear nits, and encoded PQ values on one axis without clear labels.
• Do not present a graph without preserving the script or source CSV needed to reproduce it.
• Do not keep copying a repeated plotting scaffold across scratch scripts; promote the pattern when it recurs.