Localized image and video automation: a practical pipeline
Localize generated images and videos without broken layouts: text expansion budgets, font coverage, RTL checks and version-pinned templates.
Localized image and video automation breaks at the layout, not at the translation. A practical pipeline for text-expansion budgets, script coverage, right-to-left reviews and version-pinned templates that survive every market.
A translated layout is not a localized one

Localized image and video automation fails in the same three places: a headline that outgrows its frame, a font with no glyph for the character you just translated, and a layout that assumes text runs left to right. None of these are translation problems — they appear only after the strings arrive.
The fix is not a bigger design system. Treat a localized creative as a template with per-locale inputs, a measured text budget, and a review step that looks at the render instead of the string table.
Three layers you localize separately
| Layer | What changes | Where it lives |
|---|---|---|
| Copy | Headline, subline, legal line, price format | Locale files or your translation tool |
| Typography | Font family per script, size, line height, letter spacing | Template typography inputs |
| Composition | Reading direction, badge position, image crop, safe areas | Template version or a separate project |
Most teams automate the first layer and hope the other two hold — until the first market that reads right to left or needs a wider character set. Keeping the layers apart makes a fix local: a size change for German does not touch the Japanese variant.
Budget for text expansion before you place a layer
Text expansion is the length difference between source and translation, and it hits short strings hardest. Published localization guidance gives a planning range against English.
| Target language | Typical expansion | Note for layout |
|---|---|---|
| German | 20–35% | Compounds push buttons and labels furthest |
| Finnish | 20–30% | Long words, few spaces to break on |
| French | 15–25% | Punctuation spacing adds width |
| Spanish, Italian | 15–20% | Usually fits a 25% reserve |
| Dutch, Polish | 15–25% | Check the longest label, not the average |
| Portuguese, Russian | 10–20% | Cyrillic needs wider glyph coverage |
| Arabic | 15–25% | Plus the direction change below |
| Japanese | −10 to +10% | Shorter in characters, often taller |
| Chinese | −20 to −10% | Fewer characters, more vertical weight |
| Korean | 0 to +15% | Mixed Hangul and Latin |
The same guidance recommends designing for up to twice the English length, which sounds excessive until a two-word button grows by 80%. Measure the longest translated string per input and size the text box to it plus a margin. Averages lie; the longest string is what ships.
A pseudo-locale that pads every string by 40% with accented characters shows the damage before the translations exist.
Fonts: coverage beats style
A template that renders perfectly in English can produce empty rectangles in Vietnamese, Hindi or Arabic: the chosen font has no glyphs for those characters.
JsonCut stores fonts as media: WOFF2, WOFF, TTF and OTF files upload through the media API, and the workspace library keeps the same approved files available across projects — a stable font stack per script instead of a font name that resolves differently on each machine.
| Script group | What to check before publishing |
|---|---|
| Latin extended | Vietnamese diacritics, Polish and Czech glyphs, ligatures |
| Cyrillic, Greek | Complete alphabet in both weights you ship |
| Arabic, Hebrew | Joined forms, vowel marks where used, direction support |
| CJK | Glyph set for the characters used, not just the sample |
| Devanagari and other Indic | Conjuncts and matras in the specific font |
CJK also changes vertical rhythm: the same point size often needs more line height than Latin. If your template exposes line height, that becomes a value per locale instead of a redesign.
Bind typography per locale
In JsonCut a published template exposes typed inputs, and typography is part of the bindable property groups: font size, color, background, line height and letter spacing can each be driven by a variable. That turns "German needs 92% of the headline size" into data instead of a manual edit.
| Input | Type | German value | Japanese value |
|---|---|---|---|
headline | string | "Jetzt erhältlich" | "発売中" |
headlineSize | number | 58 | 64 |
headlineLineHeight | number | 1.05 | 1.25 |
headlineTracking | number | 0 | 0.02 |
showBadge | boolean | true | false |
Number inputs can carry a minimum, maximum, step and unit, so a locale file cannot push a size into unreadable territory without failing validation. Discover the exact contract with GET /api/v2/templates/{id}/inputs first — the same discipline as in template input design.
| Approach | Works when | Trade-off |
|---|---|---|
| One template, one value set per locale | Same composition, copy inside budget | One version to maintain; a locale that needs a different composition breaks it |
| One template version per locale | RTL, different typographic scale, market imagery | Parallel versions to approve and pin |
| Separate project per locale | Structural differences | Most control, most maintenance |
Publishing captures a project version, and a later edit does not silently mutate a published template. A German batch rendered in November keeps rendering the design approved in November — the same governance argument as in brand consistency for automated production.
Right-to-left is a layout decision
Arabic and Hebrew do not just flip the paragraph alignment. The Unicode Bidirectional Algorithm defines how characters of different directionality are ordered within a line, including numbers and paired brackets. That is the character level. It does not decide which side your logo sits on, which way a progress arrow points, or where a badge belongs.
| Element | Common failure | Localized decision |
|---|---|---|
| Headline and body | Text aligned right but margins unchanged | Mirror the padding, keep numerals LTR |
| Arrows, chevrons, progress bars | Point the wrong way | Mirror the icon, not the whole image |
| Phone numbers, prices, URLs | Reordered or split | Isolate them; keep them LTR |
| Logos and badges | Move into the text column | Define a mirrored safe area |
| Video captions | Burned-in left-aligned text | Reposition per direction |
Test with real strings: a mirrored mock-up in English will not reveal how a mixed Arabic-and-number line resolves.
One batch per locale, pinned to one version
Localization multiplies rows. JsonCut bulk rendering accepts CSV or JSON, validates every row before queueing, and reports invalid rows without blocking the valid ones — the failure mode covered in why one bad row should not stop your bulk render.
| Plan | Rows per batch |
|---|---|
| Hobby | 50 |
| Pro | 250 |
| Business | 500 |
A CSV body can be up to 2,000,000 characters, and every batch is pinned to one immutable template version. Give each row an explicit idempotency key that includes the locale, such as sku-1042-de, and keep one batch per locale so a failed German run does not requeue the French one.
Media retention belongs in the plan. Bulk inputs can be temporary, which suits one-off local product shots — but 48 hours is shorter than a campaign cycle. Product images per market belong in project or persistent storage, so a re-render next week still finds them.
A QA gate that catches what spellcheck misses
A localized render needs a look at the output, not only at the string table.
| Check | Pass condition | Catches |
|---|---|---|
| Pseudo-locale | 40% padded strings render without clipping | Truncation, fixed boxes |
| Glyph audit | Longest string shows no fallback boxes | Missing characters |
| Expansion pass | Longest real translation fits its frame | Overflow, overlap |
| Bidi pass | Numbers, brackets and arrows read correctly | Reordered text, wrong icons |
| Native review | A native speaker signs the final render | Tone, false friends, formats |
Keep the reviewed render next to the template version, not in a chat thread. When someone asks why the Italian banner used 44 px instead of 48 px, the answer should be a version and a value.
Nothing here translates for you
This pipeline does not translate. JsonCut renders the strings, fonts and values you send; translation, terminology and legal review stay with your localization tooling or agency. A renderer should be predictable with your copy, not creative with it.
What it can do is make localization visible: one version per market, typography driven by data, output you can inspect per locale. The alternative is a spreadsheet of translations and a designer clicking through fourteen languages the night before launch.
FAQ
Do we need a separate template for every language?
No. One template with per-locale value sets works while the composition stays the same. Add a version per locale when direction, typographic scale or imagery structurally differ.
How do we handle text expansion in generated images?
Measure the longest translated string per input, reserve space for it plus a margin, and run a 40% pseudo-locale before the real translations arrive. Bind size and line height as numbers so each locale adjusts without a redesign.
Can Arabic and Hebrew be rendered automatically?
The bidi algorithm orders characters correctly, but layout mirroring is a design decision. Produce a mirrored version, test mixed numeral strings, and have a native speaker review the render.
Where do fonts live in JsonCut?
Fonts upload as media in WOFF2, WOFF, TTF or OTF format. Keeping them in the workspace library makes them available across projects.
Sources
https://jsoncut.com/docs/media/ — JsonCut media uploads: font formats, retention modes, workspace library — accessed 2026-11-21.
https://jsoncut.com/docs/create/variables-and-templates/ — typed inputs, bindable typography properties, version pinning — accessed 2026-11-21.
https://jsoncut.com/docs/templates/ — bulk rendering: rows per plan, CSV limits, per-row validation, immutable template versions — accessed 2026-11-21.
https://www.unicode.org/reports/tr9/ — Unicode Standard Annex #9, Unicode Bidirectional Algorithm (Unicode 18.0.0, 2026-09-01) — accessed 2026-11-21.
https://simplelocalize.io/blog/posts/text-expansion-ui-localization/ — text expansion ranges per language and the "twice the English length" guideline — accessed 2026-11-21.