— Best practices

Get the cleanest screenshot translations on the first pass.

A reference for picking models, designing source screenshots, handling expansive and compact languages, writing custom instructions, and exporting at the right device size. Grounded in how lokal actually renders — no guesswork.

— §01 · Model choice

Pick the right model for the job.

lokal supports four image models, two from OpenAI and two from Google. The trade-off is quality versus credit cost — and the cost is set per render, not per locale.

GPT Image 2

OpenAI

1cr low · 2cr medium · 3cr high

Strongest on dense, info-heavy UI — long lists, tabular layouts, multi-column dashboards. The high preset is the most pixel-faithful option lokal can render.

Pick when: Hero screenshots; dense interfaces; anywhere small mistakes will be noticed.

GPT Image 1.5

OpenAI

1cr low · 2cr medium · 3cr high

Same cost ladder as GPT Image 2, slightly less faithful on edge cases. Good baseline for first-pass batches and for screenshots that are mostly typography on a flat background.

Pick when: First-pass batches; simple typographic screenshots.

Nano Banana Pro

Google Gemini

2 credits · fixed quality

Fixed-quality. Excellent at preserving illustration style and brand color. Two credits per frame, no quality dial.

Pick when: Illustrative or photo-rich screenshots; brand-driven hero shots.

Nano Banana 2

Google Gemini

1 credit · fixed quality

Cheapest render in the catalog at one credit per frame. Quick and good enough for previews, draft locales, and screenshots you have not pinned a model on yet.

Pick when: Previews; draft passes; screenshots where price beats fidelity.

— §02 · Source design

Design the source so it can be translated.

The model is told to keep the layout pixel-identical — same colors, icons, illustrations, spacing, font family, font weight, alignment and overall image dimensions. The only thing that changes is the language of the text.

That contract works when the source gives the model room to breathe. Five rules:

  • Keep headlines to 4–7 words per line so expansive languages have room to grow.
  • Leave 12–16% horizontal margin around any text block.
  • Pin status bar, illustrations and photographs to areas the model can ignore — only the textual UI is re-rendered.
  • Avoid mid-word hyphenation and forced line breaks in the source — both are explicitly disallowed in the model prompt and will be rephrased.
  • Brand names, product names, numbers, currency symbols and version strings stay untranslated; do not stylize them so heavily that the model treats them as decoration.
— §03 · Languages

Some languages expand. Some compact.

lokal applies different rephrasing rules per language group. Knowing which group you are translating into is the single biggest predictor of how clean the first pass will be.

Expansive

+15–35% vs EN

These ten languages run longer than English at the same font size. The default behaviour is to pick the shortest natural phrasing, drop filler words and break compound words rather than overflow a container. German is the worst offender — expect occasional re-rolls on tight buttons.

  • deGerman
  • frFrench
  • esSpanish
  • itItalian
  • ptPortuguese
  • nlDutch
  • ruRussian
  • plPolish
  • trTurkish
  • ukUkrainian

Compact

≤ EN width

These five languages are typically equal-to-shorter than English in pixel width. The default behaviour is to translate fully — no compression, no abbreviation. Overflow is rare; the main thing to watch is that your source font has full CJK glyph coverage.

  • jaJapanese
  • koKorean
  • zhChinese (Simplified)
  • thThai
  • viVietnamese

The remaining locales — English, Arabic, Hindi, Swedish, Indonesian — fall into a neutral group: translate naturally and concisely, do not invent content, keep each string inside its bounding box.

— §04 · Custom instructions

Steer the model with one short line.

Every re-roll accepts a custom instruction. It stacks on top of the default rules — it cannot override the two hard rules (“translate every string”, “do not break the layout”), but it overrides every other default when in conflict.

  • Translation reads too formal for a consumer app

    Use informal address (tu / du / 너) and casual phrasing.

  • Translation drifts away from your terminology

    Prefer the term "workspace" over "project"; keep the word "app" untranslated.

  • German / Russian copy is overflowing buttons

    Aim for the shortest natural phrasing; break compound words if needed; never let text overflow.

  • Japanese honorific level is off for the brand voice

    Use ですます-form, friendly but not overly polite. No 敬語.

Keep instructions specific and short. A one-line directive (“use du-form”, “keep the word ‘app’ untranslated”) lands more reliably than a paragraph describing the brand voice.

— §05 · Export

Export at the right device size.

Default export keeps the source resolution untouched and ships a per-language ZIP with your original filenames. For the App Store, four device presets resize the output to the dimensions Apple expects.

iPhone 6.7"1284 × 2778
iPhone 6.5"1242 × 2688
iPad 13"2064 × 2752
iPad 12.9"2048 × 2732

The export ZIP is named <lang>-translations.zip. Inside, every frame keeps the filename you uploaded — drop the archive straight into App Store Connect without renaming.

— §06 · Credit math

Mix models to cut credit spend.

A two-pass strategy keeps cost predictable without compromising hero quality. Example: 10 screenshots × 20 languages = 200 renders.

  • Pass 1 — draft on Nano Banana 2

    200 renders × 1 credit each. Quick, cheap, good enough to review every frame side-by-side.

    200 credits

  • Pass 2 — promote ~20% to GPT Image 2 medium

    The frames Apple is going to scrutinize: the first two screenshots in each locale. ~40 renders × 2 credits.

    + 80 credits

  • Total

    Compared to running everything on GPT Image 2 high (600 credits), the two-pass strategy is roughly 53% cheaper — with the same hero quality where it matters.

    280 credits

Open the workspace and try the two-pass flow.

Cheapest model first, then promote the hero frames. The math is on this page; the muscle memory comes after one project.