Sepia is a header-mostly C++17 plotting library built from scratch to produce clean, publication-quality line plots at any data scale - from a handful of points to over a billion - without relying on a single external library.

The rendering pipeline is fully custom: Xiaolin Wu anti-aliased line drawing, alpha-blended RGBA pixel operations, automatic "nice" tick mark generation, and a self-contained 5x7 bitmap font - no FreeType, no libpng, no graphics framework of any kind. Output is written as PPM, a trivially simple format that converts cleanly to PNG downstream.

The centerpiece algorithmic feature is LTTB (Largest-Triangle-Three-Buckets) decimation. Rather than rendering every raw sample, Sepia downsamples massive datasets to a configurable target resolution while preserving the visual shape of the curve. At one trillion points the difference is stark: 952ms with LTTB versus 23,490ms without - a 24.7x speedup - with no perceptible loss in fidelity.

Memory is treated as a first-class concern throughout. Pixel buffers and data series use 64-byte cache-line aligned allocations (posix_memalign / _aligned_malloc) to keep hot data SIMD-friendly. Three data ownership modes are provided - owning copy, owning move, and zero-copy external reference - so callers can avoid unnecessary allocations when feeding in large existing arrays.

The API exposes a fluent PlotCommand builder returned by figure.plot(), enabling chainable style configuration (.color(), .width(), .alpha(), .label(), .marker(), .fill()). Alternatively, pre-built aggregate style structs (DataStyle, GridStyle, AxisStyle, LegendStyle) can be passed directly for batch configuration. The result is an interface that scales from a 10-line quick plot to a fully styled multi-series figure.

The library compiles with GCC 7+, Clang 5+, and MSVC 2017+ and ships with runnable examples (basic sine wave, multi-series chart, 10M-point large dataset) and a stress test suite that benchmarks render latency across sizes from 100 points to one billion.