Can optical kerning work in Typst?

What optical kerning is

Kerning changes the horizontal space between neighboring shaped glyphs. In metric kerning, the layout engine follows spacing and positioning data supplied by the font and shaping system. In optical kerning, the actual glyph shapes are also examined so visually uneven gaps can be corrected.

Large poster and presentation titles expose these problems quickly: a gap such as W|A, A|V, o|T, or 1|0 can look too open or too tight even when the numeric advances are valid. Manual tracking fixes one line, but it does not provide a reusable document-level behavior.

Context: Practical Typography's critical comparison of metric and optical spacing. This benchmark follows its central caution: optical is an option for a problem case, not a replacement for good font metrics.

Why InDesign is foregrounded

InDesign Optical is the external publishing reference in this workbench because designers already know it and use it in production. The goal is not to reverse-engineer Adobe or declare its output to be ground truth. It gives reviewers a real, inspectable answer to “compared with what?”

Every optical comparison first passes a metric-parity gate: Typst and InDesign must use the same static font instance, text, size, ligature setting, and closely matching metric output. Only then is the InDesign Optical result compared with the Typst candidate.

Inspect the full method in the InDesign baseline specification and metric-parity gate. The separate academic display report contains both 80 pt and 100 pt results.

What the algorithm does

The result of this work is a guarded hybrid algorithm, not one spacing constant. The benchmark first tested nearest-contour distance, whitespace profiles, area balance, metric-prior blending, and conservative fallback behavior. Their failures led to the full guarded-profile-hybrid research reference and then to the smaller compact-guarded runtime candidate.

1. Shape first. Work with the real glyph run and active OpenType features.
2. Keep a metric prior. Start from shaped font positioning, prevent sign changes, and bound optical movement around nonzero values.
3. Read outlines. Build font-local whitespace profiles from glyph contours.
4. Guard corrections. Reject moves that risk collisions, broken apertures, punctuation errors, or poor word rhythm.
5. Emit small deltas. Return deterministic, cacheable adjustments in em.

Read the algorithm and candidate-selection notes, the runtime and compiler prototype results, or the technical evaluation.

Measured result

The first table records the full research reference. The smaller compiler-facing candidate and its preservation trade-off follow directly below it.

These scores measure rendered position and width differences against InDesign Optical. They make regressions visible; they do not replace visual review or prove that InDesign is always preferable.

The extracted compact-guarded kernel was compared with two deliberately simpler controls on the same 30 no-ligature cases: nearest contour scored 0.2914em mean, fallback only 0.1061em, and the current bounded compact candidate 0.0222em. On the 31-case ligature suite, bounded compact scores 0.0156em mean and 0.0360em worst. The separate compiler prototype matches the workbench's expected width correction in all 61 current cases within 0.000049em.

What changed after the font-metric audit

A second audit compared the candidate with effective kerning already present in 15 Google Fonts. It retained 15,271 pairs and initially found 659 sign changes and a 0.2720em maximum difference. The generic bounded-prior rule reduces the mean difference from 0.0206em to 0.0095em, caps the maximum at 0.0300em, and removes all sign changes. No font or glyph exceptions were added.

See the current findings and known residual differences and the machine-readable font metric agreement audit and the before/after preservation report.

From workbench to a possible Typst PR

There is an active upstream discussion in Typst issue #8514, but no assigned implementation or accepted API shape. The current repository is not a merge-ready patch. Its job is to make the algorithmic direction and its tradeoffs concrete.

The exact API remains open. A deliberately small sketch is to preserve the current boolean behavior and add an optical mode rather than prematurely renaming every metric source:

#set text(kerning: true)      // current behavior
#set text(kerning: false)     // current disabled behavior

// Possible future direction, subject to upstream design:
#set text(kerning: "optical")

The technical milestones, open decisions, and useful community contributions are listed in Path to a Typst Prototype. Concrete implementation and timing evidence is in the runtime prototype report.

Repository map

Open the repository Reproduce the workbench