[Part 3] How I Built a Multilingual AI Webtoon Workflow

Building a One-Person AI Content Studio for $20/Month

From speech-bubble failures to motion comics, five-language localization, and platform-ready packages

After producing YouTube Shorts, I asked a new question:

Could AI help one creator make a webtoon and motion comic, localize the story into multiple languages, and publish it on platforms around the world? How much of that workflow could become repeatable?

To test it, I created an original office story titled Learning to Take My Place. The story follows a company where an AI employee learns the work of 27 human employees. I produced static webtoon episodes and vertical motion comics, then localized them into Korean, English, Spanish, Traditional Chinese, and Japanese.

The short answer is yes: one person can expand an idea into a script, images, dialogue, audio, localized editions, and platform packages with AI agents. But it did not work as a one-click system. Across Episodes 1–3, I turned repeated human corrections into production rules and reusable checks.

Episode 3, “The 98.4% Wrong Answer.” Its production system was shaped by the speech-bubble failures of Episode 1 and the audio failures of Episode 2.

One story, five languages, multiple platforms

The Korean master of Episode 3 contained 29 cuts. That master became 145 localized images across five languages, language-specific motion comics, thumbnails, covers, descriptions, and upload packages.

When Korean dialogue changed, the system could identify affected cuts, recalculate text layout, and rebuild only the necessary localized images and video components rather than recreating everything.

Episode	Production result	Main problem discovered	Rule carried forward
EP1	31 panels; 19 dialogue units reviewed	Speech bubbles covered faces and hands; mobile readability was weak; profile fields were missed	Protect faces and key actions; place bubbles near the speaker; check bounds and readability
EP2	32-cut webtoon and 63-second motion comic	Seven bubble-repair passes, four visual-QA failures, and six audio test/revision stages	Separate video, effects, and music; compare music candidates; check ducking and levels
EP3	29 Korean cuts; 145 images across five languages plus motion comics	Three selected cut regenerations, three BGM candidates, and missing localized thumbnail/cover items	Rebuild changed cuts only; recalculate each language; run semantic, pixel, hash, and packaging QA

The counts are not quality scores. They are production units taken from revision and QA records. Their value is that they make the rework visible and reusable.

Episode 1: when speech bubbles covered faces

The first major failure was not the story. It was the speech bubbles. Expanding a bubble to fit text covered faces and hands. Moving it into empty space created awkward tails or made the speaker unclear. Desktop text became too small on a phone, and a static layout could collide with platform UI in a vertical video.

I introduced rules to protect faces, mouths, hands, and important actions; calculate bubble size from dialogue; search for valid positions; and keep the speaker relationship readable.

Three representative cases from the 19 dialogue units reviewed in Episode 1.
Episode 2: adding music was easier than choosing it
A motion comic containing BGM and sound effects was not necessarily a good video. A track could pass technical checks and still feel too loud, weak, slow, or emotionally wrong.

I separated the visual track, sound effects, and BGM. Multiple music candidates were applied to the same scene. Dialogue sections used music ducking, and the opening, middle, ending, and music-only sections were reviewed separately.

Episode 2, Cut 16. The composition and spacing were revised so the characters and dialogue were easier to read.

The 28–37 second section in two Episode 2 revisions. The audio lineage was preserved while only the target visual section changed.

Episode 3: multilingual layout was not just translation

Replacing Korean text with a translation was not enough. English and Spanish often required more space. Traditional Chinese and Japanese still needed different line breaks, font choices, and visual balance.

For each language, the system recalculated the speech-bubble and narration-box size, line breaks, font size, and position. Automated checks looked for clipped text and elements outside the canvas. Human review checked tone, character voice, and natural language.

The same Episode 3 cut in Korean, English, and Japanese. The layout was recalculated for sentence length, line breaks, and font behavior.

The observed preparation interval fell from seven days to two
Episode 1 and Episode 2 were published seven days apart. After Episode 2 was published, the complete Episode 3 upload package passed its release audit in two days.

Manual correction in Episode 1, partial workflow rules in Episode 2, and an integrated pipeline in Episode 3.

This does not mean pure labor time fell by 71%. Start and finish times were not measured consistently. The comparable number is only the observed interval between a public release and the next audited package.

What the AI agents and vibe coding actually did

Vibe coding did not end with asking for a visual change in natural language. The reason for each correction became a layout, protected-area, audio, rights, metadata, or QA rule that an AI agent could execute again.

AI-agent execution	Human decision
Structure scenes, speakers, and dialogue; flag ambiguity	Story, emotion, character voice, and the episode hook
Calculate bubble size and positions; detect collisions	Exceptions, reading order, and protected faces or actions
Draft translations and recalculate layouts	Natural language, cultural context, and character tone
Place BGM/effect candidates and check levels	Emotional fit, usage rights, and final listening
Generate platform files, covers, descriptions, and checklists	Account access, preview, publication decision, and timing

The current system is semi-automated, not autonomous. AI agents handle repeatable production and technical checks. A person still decides meaning, emotion, rights, and publication.

Watch and read the existing English editions

No new English promotional Short or dubbed long-form video was produced for this Blogger launch. These are existing published English story editions.

Conclusion

It was possible to build, localize, and publish a webtoon and motion comic from a one-person production environment. The essential improvement was not eliminating human intervention. It was turning one human judgment into a rule that could be reused in the next episode.

If a creator has a story and can review AI output critically, a professional production team is no longer the only way to begin publishing a webtoon for readers in multiple languages.

Previous: Part 2 — How I Made 3 YouTube Shorts with ChatGPT and AI Agents

This article is an English localization of a production record first published in Korean. Korean original / 한국어 원문.

Search This Blog

Kihong Story Lab | AI One-Person Content Studio