› aurix/case studies/image-localization
Case Study · AI Image Localization/03

Image localisation, treated as infrastructure.

PixelLingua is a full-stack image localisation platform. OCR extracts the text. AI removes it. AI translates it. The result is re-rendered in the target language — all under a canvas-based annotation flow that lets vendors and reviewers score, edit, and approve every change with a full audit trail.

scroll
/metric.01
4-stage
AI image pipeline · OCR, inpaint, translate, render
/metric.02
3
Ingestion sources · ZIP, Google Drive, AWS S3
/metric.03
100%
Audit-traceable on every vendor edit
/metric.04
Canvas
WYSIWYG vendor annotation, not spreadsheets

› the challenge / 01

Translated text on a website is one thing. Translated text inside an image is another.

Strings on a website are infrastructure — they live in a database, move through a build pipeline, deploy on a cadence. Strings inside an image are not. They are baked into pixels. They have typography, drop shadows, colour treatments, and layout that the original designer chose deliberately. Translating them means undoing that work and redoing it — once for every locale.

Manual image localisation is a designer-and-translator collaboration that is notoriously slow. The designer extracts text by hand (or, if they are lucky, has the original layered file). The translator translates. The designer rebuilds the layout in the target language — accounting for string length, typography, and locale-specific characters. Each round costs hours per image. Each language doubles the cost.

For an agency localising marketing visuals, product packaging, infographics, screenshots, or in-product imagery for a global launch, this scales catastrophically. A 200-image campaign in eight languages is sixteen hundred manually-rebuilt images. There is no way to throw more designers at the problem and stay profitable.

“OCR the text. Inpaint it out. Translate it. Re-render it in target. Then put a vendor in front of a canvas to fix what’s wrong.”
— the design constraint

The challenge was to make image localisation a pipeline — fully automated by default, with a structured review layer where vendors and reviewers could correct what the AI got wrong, all on a canvas where the visual context is preserved. And to make the whole thing accept assets from wherever the customer happened to keep them: a ZIP file, a Google Drive folder, or an S3 bucket.

› the problem / 02

Two ways to localise an image. Both fall apart at scale.

The existing approaches to image localisation collapse into two shapes — a slow manual workflow, or an opaque automated one. Neither is what an agency operating at production volume actually needs.

Failure mode 01

Manual rebuild in Photoshop, per language.

A designer opens the source file, copies the text, hands it to a translator, waits, pastes the translation back, fixes the layout for the new string length, exports. Then does it for the next language. Then the next image. A 200-image campaign across eight languages becomes sixteen hundred rebuilds — a project nobody wants to staff.

throughputper image →
observed
1600 manual rebuilds
Failure mode 02

OCR-and-translate scripts with no review layer.

The pipeline runs end-to-end, but there is no way for a vendor to see the rendered output and fix what is wrong. Errors propagate silently. Glossary terms drift across files. There is no audit trail when a customer asks why a specific phrase was translated a particular way. The output ships with mistakes nobody caught.

throughputper image →
observed
Silent error propagation
/summary

Neither produces what an agency at scale actually needs: automated processing on every image, with a canvas-based review surface where vendors can correct the AI, glossary support to keep terminology consistent, and a full audit trail that survives a customer challenge.

› the solution / 03

A four-stage AI pipeline. A canvas review surface. Multi-source in, multi-format out.

PixelLingua treats image localisation as a pipeline that runs by default. AI handles OCR, text removal, translation, and re-rendering on every ingested image. Vendors review the output on a canvas — drawing, editing, scoring directly on the rendered result. Glossary and audit trail run underneath everything.

pixellingua.pipeline / live
4 stages · raster in · raster out
A1OCRHELLOEXTRACT TEXTA2InpaintcleanedREMOVE TEXTA3Translateen → jaこんにちはglossary awareWITH GLOSSARYA4Renderこんにちは✓ deliveredTARGET LOCALE▸ asset enters as raster · exits as localised raster · audit trail preserved
01

Four-stage AI image pipeline

OCR → inpaint → translate → render. Every image runs the full pipeline by default. The output is a target-language image, not a translation file the customer has to assemble themselves.

02

Canvas-based vendor annotation

Vendors review on the rendered image, not a list of strings. Drawing, editing, and scoring happen in visual context — so a typography issue is fixed where it lives, not extracted into a separate ticket.

03

Multi-source ingestion + multi-format export

Customers send assets however they keep them — ZIP, Google Drive, AWS S3. The platform ingests all three. Localised output exports back to the same set, so the workflow drops into existing asset management.

04

Audit trail + glossary support

Every vendor edit, every AI decision, every glossary substitution is logged per asset. Glossary entries enforce term consistency across thousands of images. When a customer asks why a phrase was rendered a particular way, the answer is one query away.

› the result / 04

Image localisation runs as a pipeline, with humans only where they add value.

PixelLingua is operating as a full-stack image localisation platform. AI handles routine processing. Vendors review on a canvas. The customer gets localised images back through the same channel they sent them in.

/result.01
200×
Throughput vs. Photoshop manual rebuild per locale
/result.02
3-in / 3-out
Source and target formats supported · ZIP, Drive, S3
/result.03
0
Manual file handling between ingestion and delivery
01

Image localisation became a pipeline, not a project.

Customers drop assets into ZIP, Drive, or S3. The pipeline ingests, runs the four AI stages, surfaces the output on a canvas for vendor review, and exports localised images back. The workflow is the same whether the batch is twenty images or two thousand.

02

Vendors got a canvas, not a spreadsheet.

Reviewers see the rendered image, the AI's translated text in place, and the controls to fix it directly on the canvas. Typography issues are corrected where they appear. There is no extract-fix-reimport loop — the surface is the image itself.

03

Audit trail preserved every change.

Every AI decision, every vendor edit, every glossary substitution is logged per asset. When a customer asks why a particular phrase was rendered a particular way, the answer is queryable — the agency does not have to reconstruct the chain from memory.

04

Multi-source ingestion absorbed any asset shape.

Customers do not change how they store images to use the platform. ZIP archives, Google Drive folders, S3 buckets — all three flow into the same ingestion. Output exports back through the same channel, so the platform fits into existing asset management rather than replacing it.

05

Glossary support kept terminology consistent across thousands of images.

The translation stage is glossary-aware. A brand-specific term is rendered the same way whether it appears in image one or image one thousand. The glossary itself is editable through the platform — corrections made by reviewers feed back into the term base for future jobs.

› built with

ReactDjangoGoogle Cloud VisionOpenAIAWS S3
› get started

Ready to make image localisation a pipeline, not a project?

Talk to us about how PixelLingua can plug into your localisation workflow. We will walk you through the platform, run a sample batch from your own assets, and tailor a deployment to the source formats and review process you already have.

30-min walkthrough · No commitment
Sample batch · Bring your own assets
Pilot-ready · ZIP, Drive, or S3 from day one