Five rewrites compete for every line.
One wins, and it has to say why.
Every dialogue line goes through a structured review process. Each adaptation is checked against the source and comes with a plain-language reason for creative changes.
Five rewrites, scored against the anchor
“You came back. I knew you would.”
Rewrite 1
Rewrite 2
Rewrite 3
Rewrite 4
“Kept the character’s guarded tone while making the line sound natural in this scene.”
Translator
Produces one literal, anchor translation of the source line. This isn't shown as the final output — it's the floor everything else is checked against, so a later rewrite can be judged against what the line actually says, not just what it might mean.
Creative Adapter
Produces five differently-angled creative rewrites of the same line, each taking a different liberty with phrasing, idiom, or rhythm — never five copies of the same idea.
Judge
Scores each of the five rewrites against the Translator's literal anchor and against each other, picks a winner, and writes down the specific, real reason it departs from a literal translation. If none of the five earns a genuine improvement, the literal anchor ships instead of a manufactured difference.
Why five rewrites and a Judge, not one pass
A fluent translation can still miss what makes a line work in a comic: the character speaking, the relationship between speakers, the surrounding artwork, and the moment in the story. Comparing alternatives against the source helps editors see those trade-offs instead of accepting the first plausible rewrite.
It's also why the Judge is a scoring model reading against a literal anchor, not an automated metric like BLEU. Callison-Burch et al.'s 2006 analysis of BLEU — the standard automated translation-quality score — found that a higher BLEU score doesn't reliably track a better human-judged translation, because BLEU rewards closeness to one reference wording, not whether a differently-worded line captures the same feeling better. Scoring against the literal anchor's actual meaning, the way the Judge does, sidesteps that specific blind spot.
Before a rewrite ever reaches the Judge, it's also checked for a specific set of failure patterns an automated pass can catch reliably: dialogue being flattened into an unnatural block of prose, or a rewrite so close to the source that it offers no meaningful improvement. A rewrite that trips one of these checks is sent back for a fresh attempt before it's ever put in front of the Judge, rather than being scored as-is.
Webtoons & comics
OCR, a cast profile, then the identical Writers' Room.
A chapter's panels go through Google Cloud Vision OCR, which detects each speech bubble's text and the order it's meant to be read in. Before any line is translated, a separate pass reads the whole chapter and builds its Chapter DNA — a profile of who's speaking, how each character speaks, and the tone of the scene — so a character's established voice carries into the per-bubble prompt for every one of their lines, not just the first. From there, every line runs through the same structured review process — with a clear source reference and a reason for each creative change.
An optional redraw pass can erase the original lettering, inpaint the artwork underneath it, and typeset the adapted line back into the bubble — with overlapping speech bubbles merged into a single region before typesetting so text doesn't collide, a font picked per region or set as a page default from a bundled set of open-license comic lettering fonts, and a bounded retry if the inpainting service call fails rather than leaving a panel half-processed.
Panel upload
One image per panel, in reading order
OCR
Google Cloud Vision reads each bubble's text and position
Chapter DNA
Genre, tone, and cast profile for the whole chapter
Writers' Room
The same Translator → Adapter → Judge pipeline as music
Adapted script
Literal, adapted line, and why — per panel
Try it on a real result first — see the worked example or read the FAQ.
Sources
- Wu et al., "Google's Neural Machine Translation System" (2016), Google Research / arXiv — what a single-pass neural MT system is architecturally optimized to do
- Callison-Burch, Osborne & Koehn, "Re-evaluating the Role of BLEU in Machine Translation Research" (EACL 2006) — why an automated closeness score doesn't reliably track human-judged quality - the reason a Judge stage exists instead of a metric