Why AI Breaks Korean Text in Ads — and Why the Real Fix Is Preservation
ImageFactory Engineering · Published 2026-06-12
The reason AI garbles Korean text in ad images is simple: generative models draw letters as visual patterns, not as language. And Korean is structurally disadvantaged — its alphabet composes into 11,172 distinct syllable glyphs, versus the 26 shapes of the Latin alphabet a model has seen billions of times. Better prompting lowers the failure rate but never reaches zero. So for ad work — where brand names, prices and legal copy must be exact — the real fix is not generating text better, it is never redrawing the text at all: preserve the original text pixels and let AI rebuild only the background and layout. This post covers the mechanics, where the known fixes fall short, and how preservation works in practice.
AI draws letters as pictures
Ask DALL·E, Midjourney or Stable Diffusion to "write SALE" and the model does not understand the string S-A-L-E — it imitates pixel patterns that look like SALE from its training data. UX Collective's analysis explains this well: to a model, text is not symbolic meaning but a cluster of co-occurring visual elements. It draws something letter-like, with strokes missing or glyphs that don't exist.
Why Korean breaks harder than English
English has 26 letter shapes (52 with case), and Latin script dominates training data. Korean is different:
- Composed syllables: the Unicode Hangul syllable block contains 11,172 characters. Glyphs with final consonants like "값" are each a separate pattern to the model.
- Training frequency: Korean image-text pairs are scarce in global training data relative to English. Seen less, wrong more.
- Fake syllables: models recombine plausible strokes into "Korean-looking characters that don't exist". A non-speaker might not notice; a Korean user spots it in half a second — and your ad's credibility collapses in that half second.
The four known fixes, and their limits
| Fix | Effect | Limit for ad work |
|---|---|---|
| Prompt tuning (quoted phrases, short words, typography hints) | Better odds | Lowers the odds, never guarantees. A 1% typo rate on prices or brand names is fatal |
| Switching to text-strong models (Ideogram, Nano Banana Pro, etc.) | Big improvement | Latest models are far better at Korean, but long copy, small type and logo faces remain unstable |
| Writing in English instead | Avoids the issue | Not a real solution for ads targeting Korea |
| Generate without text, overlay in an editor | Most reliable | Fine for one size. Thirty placements means thirty manual passes — automation collapses |
That last row is the crux. Ads never ship as a single image — the same creative gets adapted into dozens of placement specs (Meta feed, IG Story, Google Display, Kakao Bizboard…), and regenerating text per size multiplies the chances of breakage.
So here's what to do — a working checklist
When you must generate text (first version of a new creative):
- Keep copy to 1–3 words, and put the exact phrase in quotes in your prompt.
- Use a current model with proven Korean rendering (Gemini Nano Banana family, etc.), generate 2–3 candidates, and pick a clean one.
- Exclude must-be-exact text — prices, brand names, legal lines — from generation entirely. Generate a text-free image and overlay them in an editor.
At the size-adaptation stage:
- Don't regenerate text per size. Adapt from one approved master with text preserved — thirty sizes should mean one preservation pass plus background rebuilds, not thirty regenerations.
- For placements with big ratio changes (like 9:16 Story), check that text hasn't been pushed outside the safe zone (per-placement safe-zone values are free in the ad size guide).
At review:
- Have a Korean speaker review glyph by glyph, zoomed in — a single stroke or final-consonant difference is invisible at a glance.
- Review order: brand name → numbers (price, discount) → legal copy → body copy. Ordered by the cost of being wrong.
How ImageFactory solves this
ImageFactory banner resize automates steps 4–5 above. Here's how it works:
- Upload one master creative — the approved version with the Korean copy in it.
- Pick target placements from 1,400+ specs, or bulk-load custom sizes from a spreadsheet.
- AI rebuilds only the background and layout. Text, logos and product shots keep their original pixels — the chance of garbled Korean isn't "low", the path to it doesn't exist structurally. On big ratio changes, preserved elements are repositioned inside the safe zone.
- You receive outputs with per-platform safe zones, file formats and size limits already applied — dozens of sizes in about ten minutes.
Because the principle is preservation, it is language-agnostic: it behaves identically across 15 languages, from Korean to Japanese to Arabic, and the same engine runs inside the Figma and Photoshop plugins. The fastest way to verify is the 14-day free trial with your own creative.