Text rendering on a GPU is one of those problems that looks trivial until you try it. Every scalable font stores glyphs as vector outlines — quadratic or cubic Bézier curves filled by a winding rule. CPU rasterization is solved. GPU rasterization, where you need thousands of glyphs per frame at arbitrary scales and transforms without re-baking anything, is where every method makes its compromise. This piece walks the full ladder from worst to best, and it does so clearly enough that you could make an engineering decision at the end. The tour starts at the bottom: the texture atlas. Rasterize each glyph once into a shared texture, draw quads. Fast, portable, runs on anything. But scale past the baked size and you get blur; shrink and you get shimmer. CJK glyph sets at multiple sizes are a memory disaster. The atlas is a solved problem for constrained use cases and an unsolved one for everything else. Signed distance fields (SDF), introduced by Valve's Chris Green in 2007, store distance-to-edge instead of pixels. The distance field interpolates smoothly, so magnification stays clean and antialiasing is cheap. But bilinear interpolation lies about corners — the point of an 'A', the notch of a 'K' — because sharp corners are discontinuities in the field. Viktor Chlumsky's multi-channel SDF (MSDF) fixes this by encoding three distance channels whose median reconstructs corners. MSDF is the current sweet spot for teams willing to bake an atlas: MIT-licensed msdfgen, sharp at high magnification, but still fundamentally a baked texture with all the costs that implies. Tessellation methods — Loop-Blinn, NVpathrendering, Pathfinder, Rive — skip textures entirely and turn outlines into GPU geometry. Rive's renderer, open-sourced in 2024, hits 120 fps on animated vector art. These are genuinely resolution-independent and strong for animated vector graphics, but they pay in tessellation cost, geometry blowup for complex glyphs, and sometimes vendor-specific hardware dependencies. Slug, published by Eric Lengyel in 2017, is the endpoint of the ladder. No atlas, no tessellation. Glyphs live as curve data in a GPU buffer with a lightweight band-based acceleration structure. The fragment shader casts a ray per pixel, finds Bézier crossings, counts winding, and computes exact coverage analytically. The key innovation — 'root eligibility' — is a precise rule for which curve-ray intersections count, eliminating cracks and double-counts at shared endpoints. The same glyph is sharp at 6 pixels or 6000, under any 3D transform including perspective, with no baking cost when text changes every frame. Lengyel patented Slug in 2019 and dedicated the patent to the public domain on March 17, 2026. That act enabled Slughorn, the C++20 implementation described here, and enabled this article to exist at all. The piece is honest about what each method costs and where each wins — atlas for constrained hardware, SDF/MSDF for the baked-texture sweet spot, tessellation for animated vector art, Slug for unconstrained dynamic text at any scale. What makes this piece work is its refusal to declare a winner and walk away. Every method has a use case. The atlas is still the right call on hardware where fragment shader complexity is expensive. MSDF is excellent when you can afford baking and don't need extreme scale ranges. Slug wins when you cannot predict how text will be viewed — free cameras, arbitrary zoom, perspective, live-updating content — because it computes coverage after the transform rather than before. The piece gives you the decision tree, not the sales pitch.
