Static Ad Localization Requires Copy-Only Edits and Language-Specific QA
ElevenLabs argues that static-ad localization can be automated without rebuilding creative, provided the model is tightly instructed to alter only customer-facing copy and preserve logos, brand names, typography, layout and imagery. Its tutorial shows how to build a reusable flow that creates language-specific versions from a single input ad, then uses an LLM check to flag translation and text-fit errors. For larger paid-media operations, the company positions its Ads Engine as the enterprise option for translating and distributing static and video creative across up to 90 languages.
Reliable localization starts by limiting what the model may change
A localized ad is not simply a translated image. The operational requirement is narrower: replace customer-facing copy while preserving the asset’s identity—its logo, brand and product names, photography, background, typography, composition, and button treatment.
ElevenLabs’s flow begins with the original static ad as an image input connected to an Image Generation node. In the demonstrated Einstein creative, “The Sound of Genius” and “Einstein, now on ElevenLabs” are translated, while the ElevenLabs logo remains untouched.
The prompt is the control layer. It should explicitly name every text element that may change: headline, subheading, body copy, call-to-action button, badges, prices, and offer labels. It should separately prohibit translation of brand names, product names, and anything inside a logo.
Just as important, it needs to specify the preservation rules. The model is told to keep font style, weight, color, size, position, alignment, outlines, and effects; not to move, restyle, or regenerate the product, photography, background, logo, button, or any other element; and to change only words.
Copy fit is part of that instruction, not an afterthought. The translation should use phrasing a local copywriter would naturally write rather than literal word-for-word language, while remaining roughly the original line length. If a translated phrase runs long, it should be reduced slightly within its existing text box rather than overflowing or overlapping another element. Calls to action must still fit within their buttons and retain their padding. Prices, numbers, and currency symbols remain unchanged.
The resulting French creative renders the headline as “Le Son du Génie” and the subheading as “Einstein, désormais sur ElevenLabs,” while retaining the ElevenLabs logo and the rest of the original composition.
Choose the generation settings for text fidelity, not maximum output
ElevenLabs suggests two image-generation models for this kind of controlled edit: Nano Banana 2 and GPT Image 2. Both are presented as high-quality, precise image-editing options with output up to 4K. Their tradeoff is speed against close handling of text and long instructions.
| Model | Best for | Speed and cost | Watch out |
|---|---|---|---|
| Nano Banana 2 | Fast, precise editing at full resolution | Faster than GPT Image 2; medium cost | Weaker on text and layouts |
| GPT Image 2 | Precise editing and following long prompts exactly | Very cheap at 1K; slower and expensive at 4K | Slows down at 4K high quality |
For the French translation, ElevenLabs selects GPT Image 2 because the job depends on executing a detailed preservation prompt. The aspect ratio can be set for the placement—for instance, 1:1 for a square static ad—or left on auto, which uses the reference image’s resolution.
The resolution guidance is similarly practical: most placements need at least 1K, 4K is usually overkill, and 2K can be the useful middle ground. The flow then runs at the highest available quality setting.
Reuse the translation policy, then check the output that matters
Once the preservation and translation instruction works for one language, it can be duplicated into separate branches. A French Image Generation node can be copied and its prompt changed to Spanish, German, or additional target languages. All branches retain the same source ad and the same rules about logos, typography, layout, and copy fit.
For subsequent campaigns, the operator replaces the original image input rather than rebuilding the flow. ElevenLabs says this setup can generate seven, 10, 15, or 20 language variants within seconds. The visible multi-branch output includes Spanish, German, Portuguese, Italian, and Polish headline variants from the same Einstein asset.
An LLM verification node can be attached to each translated output. It is not intended to judge the creative as a whole. Its job is to check whether the text is correctly translated and still usable in the existing asset.
The verifier reviews the headline, subheading, CTA, badges, prices, and small print. It returns “yes” and an English rendering of each text element if the ad passes; otherwise, it returns “no” and identifies the affected element. The requested checks cover source-language text left behind, mistranslation or unnatural phrasing, spelling and character errors, text that is cut off or collides with other elements, and CTAs that no longer fit their buttons.
Brand and product names, including text inside logos, remain outside the verifier’s correction scope. It is also instructed not to comment on imagery or layout except where text collides with another element.
ElevenLabs uses Gemini 3.5 Flash for this step, describing it as one of the cheaper LLM models, with thinking enabled. On the French example, the node returns “Yes” and renders “Le Son du Génie” and “Einstein, désormais sur ElevenLabs” back into English as “The Sound of Genius” and “Einstein, now on ElevenLabs.” The result gives an operator who does not speak the target language both an acceptance signal and a readable account of the generated copy.
Use the flow for configurable static production; use Ads Engine for deployment at scale
The self-serve flow is a configurable static-image workflow. An operator uploads an ad, connects it to image-generation nodes, writes the translation and preservation prompt, manually duplicates branches for each target language, and optionally attaches verification nodes. It is designed to be reused by replacing the input image and running the established branches again.
For paid-ad operations at larger scale, ElevenLabs positions Ads Engine as the enterprise alternative. It is described as handling both static and video creative, generating translations and variations of winning ads, connecting directly to an ad library, and allowing outputs to be pushed live in a few clicks. The workflow graphic associates it with delivery to Meta Ads, Google Ads, and LinkedIn Ads.
The practical distinction is not merely volume. The flow is a hands-on system for controlled static-ad translation, where language branches and QA connections are configured on the canvas. Ads Engine is presented as a scaled localization and distribution product, extending from creative translation into ad-library and platform workflows.