Vision AI game localization QA — detect UI text errors from screenshots and video.
Upload game screenshots or a video clip. Vision models identify untranslated strings, truncation, encoding errors, and terminology issues. Export a structured LQA report as CSV.
- 🖼️ Screenshot analysis — drag-and-drop PNG / JPG / WEBP
- 🎬 Video frame extraction — upload video, set interval (e.g. every 5s), frames extracted in-browser via Canvas
- 📄 Reference string CSV — upload
source,targetpairs to detect untranslated content - 🏷️ Severity classification — Critical / Major / Minor per issue
- 📋 Structured LQA report — export full CSV with frame, severity, category, description, suggestion
- ☁️ OpenAI-compatible + DashScope —
gpt-4o,qwen-vl-max, or any vision-capable model - 🔷 Anthropic Claude —
claude-3-5-sonnetrecommended for vision tasks - ⚡ Local Ollama — use
llava:13b,minicpm-v, or any local vision model
| Category | Description |
|---|---|
| Untranslated | Source language text still visible in UI |
| Truncation | Text cut off at UI boundaries |
| Overflow | Text extends outside UI element |
| Encoding/Font | Garbled or missing characters |
| Terminology | Wrong term per glossary |
| Fluency | Grammar or naturalness issues |
| Placeholder | Developer string or variable leaked |
| Reference Mismatch | Differs from provided reference string |
git clone https://github.com/Grayson0130/gamelqa.git
cd gamelqa
npm install
npm run dev
# → http://localhost:3001| Mode | Provider | Example model |
|---|---|---|
| ⚡ Local | Ollama | llava:13b, minicpm-v |
| ☁️ OpenAI fmt | OpenAI | gpt-4o |
| ☁️ OpenAI fmt | DashScope | qwen-vl-max |
| 🔷 Anthropic | Claude | claude-3-5-sonnet-20241022 |
source,target
"ATTACK","攻撃"
"MENU","メニュー"Next.js 14 · TypeScript · openai SDK · @anthropic-ai/sdk · Canvas API · papaparse
MIT © Jinhui Zhao