Game-Translator is a Windows 11 desktop application for real-time screen OCR and translation overlays in games, comics, and other bounded on-screen regions. The solution contains four production modules (Domain, Application, Infrastructure, and UI) plus one verification project (GameTranslator.Tests).
The current normal pipeline is: manual OCR zone → GPU Paddle text detection → bounded writing-system grouping → Tesseract crop recognition → the explicitly selected translator → per-region overlay. Windows OCR and Tesseract remain supported OCR engines; the detector worker is an Infrastructure adapter, not a fifth product module.
For installation, first-run setup, OCR language packs, providers, cache, hotkeys, and diagnostics, see the user guide.
The source repository deliberately excludes third-party game/manga screenshots, local OCR/translation reports, release candidates, Paddle model files, and Tesseract language data. It does contain a small tracked deterministic calibration set; generated evidence and owner-provided source material remain local-only under ignored artifacts/, work/, outputs/, application-data, and tessdata/ paths unless the owner explicitly promotes a specific artifact.
To assemble the current ADR-030 live-Paddle package class, use a Windows x64 host with:
- .NET SDK 9;
- CPython 3.12.10 x64, installed from the official Python release;
- an NVIDIA GPU and driver that make
paddlepaddle-gpuavailable; - network access for the first bootstrap only.
From a clean clone, run:
.\tools\bootstrap-paddle-runtime.ps1 `
-PythonRuntimeRoot "$env:LocalAppData\Programs\Python\Python312"
.\tools\verify-paddle-runtime.ps1 `
-RuntimeRoot .\work\paddle-runtime-win-x64
.\tools\build-track-d-opt-in-release.ps1 `
-BootstrapRuntimeRoot .\work\paddle-runtime-win-x64 `
-ValidateRuntimeOnly
dotnet restore GameTranslator.sln -r win-x64
dotnet build GameTranslator.sln -c Release --no-restore
dotnet test GameTranslator.sln -c Release --no-build --no-restore
.\tools\build-track-d-opt-in-release.ps1 `
-BootstrapRuntimeRoot .\work\paddle-runtime-win-x64 `
-TesseractLanguagePacks eng,jpn,jpn_vert,chi_sim,chi_sim_vert,tha,kor `
-SelfContained `
-ReleaseName v0.1.0-local-paddleThe bootstrap pins CPython/Paddle package versions (including the official Paddle CUDA 12.9 wheel index) and verifies the PP-OCRv6 detector and seven Tesseract packs by SHA-256. The first run downloads only their official distributions. The package build requires the exact locked language-pack set and verifies every pack before copying it; if the GPU runtime is unavailable or a language pack is missing, unexpected or changed, the script fails rather than silently producing an incomplete package.
Current publication boundary: source and tests are maintained on main; r61 remains a locally verified release candidate. Public binary releases require a separate release decision.
The current source includes the r61 feature set: opt-in watermark filtering (Off/Cautious/Strong), short-line grouping corrections, Bing Web session and connection recovery, consistent overlay DPI conversion, source-height retention with optional automatic font sizing (ADR-033), whole-word wrapping and final render containment.
Verification: Release build with zero warnings/errors; 835 tests (819 main plus 16 calibration); real WPF layout checks at 100/125/150/200 percent; packaged Tesseract/Paddle smoke and local r61 integrity checks passed. The owner confirmed improved Bing behavior and strong watermark filtering. Final owner verification of the r61 clipping/wrapping correction remains pending.
The latest portable is the locally verified, unpacked r61 candidate. It is not a signed public release or a GitHub Release asset. Generated screenshots, private OCR text, diagnostic logs, models, runtimes and portable binaries stay local. See the delivery history and user guide.
Start here:
- Documentation index
- User guide
- Project Constitution
- Technical Specification
- Architecture
- Implementation Roadmap
- Sprint Plan
- Definition of Done + Quality Gates
- AI Development Rules
Governance and AI-agent materials:
Developer module guides:
- Language: C#
- UI: WPF
- Architecture: Clean Architecture + MVVM
- Capture: Windows Graphics Capture
- OCR: GPU Paddle text detector → bounded grouping → Tesseract crop recognition; Windows OCR and Tesseract remain supported engines
- Translation cache: SQLite
- Secrets: Windows Credential Manager
The application must not inject into game processes, read game memory, bypass anti-cheat systems, or use DLL injection.
For packaging from a separately verified language-data directory, pass -TesseractSourceDirectory to the build script (or runtime verifier). All seven locked packs are still hash-checked. This avoids replacing any locally preserved root language data.