이미지: YouTube — 일레븐랩스 영상 갈무리
Summary
- ElevenLabs has released a workflow template within ElevenCreative Flows that automatically translates YouTube thumbnails by language
- GPT Image 2 changes only the headline text while preserving the layout, logo, and design, and an LLM node re-verifies translation accuracy
- The template bundles thumbnail upload and language slots as inputs with translated thumbnails as output, making it a reusable template
- 발표일
- 2026년 8월 19일 (일레븐랩스 유튜브 튜토리얼)
- 핵심 기능
- ElevenCreative Flows 내 썸네일 자동 번역 워크플로
- 이미지 생성 모델
- GPT Image 2 (16:9, 4K 출력)
- 검증 방식
- LLM 노드가 생성된 썸네일을 다시 읽어 번역 정확성 확인
- 입력/출력
- 썸네일 업로드·언어 슬롯 입력 → 번역된 썸네일 출력
- 배경 기능
- 유튜브의 언어별 개별 썸네일 설정 지원
Different thumbnails for different languages, but design was the bottleneck
YouTube supports a feature that lets creators assign different thumbnails by language for a single video. Spanish viewers see a thumbnail with Spanish text, French viewers see one with French text, and so on. Add translated titles, localized subtitles, and dubbed audio, and the entire video arrives in the viewer's own language. The problem was thumbnail design. Translated headlines often didn't fit neatly into the original layout, forcing creators to redo the work in a design tool every time they added a new language.
What got automated
A tutorial ElevenLabs published on August 19 shows how to replace this repetitive work with automation inside ElevenCreative Flows. After uploading the original thumbnail image once, it's connected to an image node that runs OpenAI's image generation model GPT Image 2 at a 16:9 aspect ratio and 4K resolution. That node is instructed to translate only the headline text while keeping the layout, logo, and design unchanged. Duplicating the same node and swapping only the language produces French, Spanish, German, and other versions in sequence.
An additional verification step ensures that even translations into languages the user can't read visually aren't published unchecked. An LLM node reads the generated thumbnail image back and confirms whether the translation is correct, reporting exactly what text appears. Adding a text node that turns the language itself into an input value completes a reusable template in which the LLM rewrites the prompt on the fly. The final result is delivered as a template that takes thumbnail upload and language slots as inputs and produces translated thumbnails as output.
How to use it
Where to start
This workflow was built inside ElevenLabs' ElevenCreative Flows. Along with the tutorial, ElevenLabs released the completed template, so users can grab it directly without building the nodes themselves.
Step-by-step usage
- Upload the original thumbnail image into the template.
- Connect an image node running GPT Image 2, setting the output ratio to 16:9 and resolution to 4K.
- Enter a prompt instructing it to translate only the headline text while keeping the layout, logo, and design unchanged.
- Duplicate the same node and change only the language to generate as many versions as needed.
- Add an LLM verification node that reads the generated thumbnail back and checks both translation accuracy and the actual text used.
- Turn the language into a text input value so the LLM automatically rewrites the prompt.
- Save thumbnail upload and language slots as inputs and translated thumbnails as output to complete the template.
Who can use it
The original report didn't mention any specific pricing plan or regional restriction. The finished template was presented as something that can be shared with a team or applied directly to a personal channel.
What you can try
For example, a YouTube channel targeting viewers in multiple countries could use a single original thumbnail to simultaneously generate Spanish, French, and German versions and attach them directly to region-specific uploads. Channels that have already prepared localized subtitles and dubbing could match the thumbnail to the same language, making the very first thing a viewer sees feel natural from the start.
Workflow structure
| Node | Role |
|---|---|
| Image node (GPT Image 2) | Translates only the headline text to generate a new thumbnail |
| LLM verification node | Checks translation accuracy and reports the actual text |
| Text input node | Takes language as an input value and rewrites the prompt |
Editor's view
On August 17, ElevenLabs launched an MCP connector for managing voice and chat agents inside Claude, and now it has attached an image generation model to automate thumbnail translation as well. Viewed together, the two announcements suggest the company is expanding beyond its original territory of voice synthesis into becoming a tool company that localizes an entire video, language by language. Subtitles, dubbing, and voice were already its strengths; thumbnail design was the last gap.
Previously, running a channel in multiple languages required three separate jobs: a translator to render subtitles, a voice actor to record dubbing, and a designer to rebuild the thumbnail for each language. This workflow is notable for compressing that last step into a single automated node. The verification step in particular—having an LLM read the generated thumbnail back and report the actual text—reads like a practical design choice meant to build trust even for languages the user can't visually verify.
According to domestic reports, ElevenLabs has been expanding its localization strategy since last July, including recruiting creator ambassadors targeting the Korean market. Korean YouTube creators or media companies running multilingual channels could plug a template like this into their existing subtitle and dubbing pipeline right now to experiment with cutting preparation time for language-specific uploads. That said, for thumbnails with complex logos or brand colors, it's safer to test a few languages directly and check how precisely the automated translation preserves the layout.
Similar automation is likely to appear at other companies building subtitle and dubbing tools within the coming weeks. This looks like the early stage of a shift toward workflows that produce language-specific versions of a video all at once as a standard part of the production process.
