What Korean Thumbnails Taught Us About On-Image Text
Korean on-image text breaks in ways English never does — sentences fragment on grammatical particles, and image models garble Hangul glyphs. Two hard constraints, and how a generation pipeline has to handle them.
Rendering readable text inside a thumbnail is already hard. Doing it in Korean surfaced two problems English hid from us entirely — one about grammar, one about glyphs. Solving them well is most of what separates a Korean thumbnail that looks native from one that looks broken.
Problem 1 — Korean sentences fragment where English doesn't
A thumbnail headline has to be short, so a long title gets shortened. In English you can usually cut at a space and still have something readable. Korean punishes that. The language is head-final and particle-driven, so a cut in the wrong place leaves the phrase mid-clause with no resolution.
Chop "메타 시니어가 주니어를 절대 뽑지 않는 이유" carelessly and you get "메타가 주니어를" — a subject and an object marker dangling with no verb. It's not just awkward; it's grammatically incomplete, and a native reader clocks it instantly as broken.
The endings you can't end on
There are whole classes of endings a Korean headline must never stop on, because each one promises more words that never arrive:
- Case particles — 을/를/의/에/에서/에게/까지. They mark a noun's role and demand a verb that isn't there.
- Bare conditionals — 하면/라면/려면. "코딩만 열심히 하면" states a condition and withholds the consequence.
- Relative-clause modifiers — verb stems ending in 는/한/된 ("코딩 노가다 끝내는"). They expect a head noun the headline never supplies.
The endings that work
A strong Korean headline lands on something self-standing. The reliable patterns are nominalized or sentence-final: a noun head like ~이유 / ~방법 / ~진실, an enumerated ~N가지, a flat declarative like ~합니다 / ~했다, or a native curiosity marker like ~한 썰. These read as complete on their own, which is the whole requirement — a viewer must be able to parse the line with no other text next to it.
This is why a good pipeline doesn't just cut a title to length. It re-derives a headline that is guaranteed to end on a complete thought, and checks the ending morphologically before it ships.
Problem 2 — image models garble Hangul glyphs
The second problem is rendering. Image models are trained overwhelmingly on Latin text, so Hangul often comes out with malformed or invented syllable blocks — a 자격증 that renders as 자껵증. At a glance it looks like text; on inspection it's nonsense, and it reads as low quality.
The defense is verification. After the render, a vision model transcribes the drawn text and compares it against the intended headline, flagging garbled glyphs and stray words. When it fails, the pipeline runs one targeted repair to fix the letters while preserving the composition. You can't trust an image model to draw Korean correctly on the first pass — you have to check it, every time.
The lesson
Korean forced us to treat on-image text as a verified output, not a hopeful one — complete grammar and correct glyphs, both checked, not assumed. It's a better standard for every language; Korean just made the failures impossible to ignore.