ScreenTranslate

A floating overlay that OCRs and translates any on-screen text in real time into your chosen language. Drag it anywhere — no screenshots, no copy-paste.

Perfect for untranslated visual novels, games, apps, and any on-screen text.

Features

• Frameless always-on-top window — drag via ⋮⋮ handle, resize from bottom-right
• Swappable OCR engine for different languages
• Swappable LLM backend (DeepSeek / OpenAI / Ollama / custom compatible APIs)
• Customizable target language
• Shortcuts: Ctrl+L toggle, Escape clear, F5 force refresh

Requirements

• Windows 10+
• Python 3.10+
• CUDA (optional — Florence-2 auto-uses GPU if available)

Installation

python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

OCR Model

Default engine is Florence-2-base (Microsoft lightweight VLM, 0.23B params, ~0.5 GB).

Auto-downloads on first run, or manually:

pip install huggingface_hub
huggingface-cli download microsoft/Florence-2-base --local-dir Florence-2-base

OCR accuracy varies by language. Florence-2 works well for English. For other languages (Japanese, Korean, Arabic, etc.), use an OCR model optimized for that language.

Configuration

Edit config.py in the project root:

# ── OCR Engine ──
OCR_BACKEND = "src.engine.ocr_florence.FlorenceOCREngine"

# ── LLM Backend ──
LLM_BACKEND = "deepseek"     # deepseek / openai / ollama / custom URL
LLM_API_KEY = "sk-xxxxx"     # API key (not needed for Ollama)
LLM_BASE_URL = ""            # Leave blank for default, or set custom URL
LLM_MODEL = ""               # Leave blank for default, or specify model name

DeepSeek (default)

LLM_BACKEND = "deepseek"
LLM_API_KEY = "sk-xxxxxxxxxxxxxxxxxxxxxxxx"

Get a key: platform.deepseek.com/api_keys

OpenAI

LLM_BACKEND = "openai"
LLM_API_KEY = "sk-xxxxxxxxxxxxxxxxxxxxxxxx"
LLM_MODEL = "gpt-4o-mini"

Local Model (Ollama)

No internet, completely free.

ollama pull qwen2.5:7b

LLM_BACKEND = "ollama"
LLM_MODEL = "qwen2.5:7b"

Custom OpenAI-Compatible APIs

vLLM, LocalAI, LiteLLM, one-api, etc.:

LLM_BACKEND = "http://localhost:8000/v1"
LLM_API_KEY = "not-needed"
LLM_MODEL = "qwen2.5-7b-instruct"

Custom OCR Engine

Implement a single method:

class MyEngine(OcrBackend):
    def recognize(self, image_bgr: np.ndarray) -> List[OcrResult]:
        ...

Results contain bbox (coordinates), original_text, and confidence.

Point config.py to your engine:

OCR_BACKEND = "my_package.my_ocr.MyEngine"

Dynamically loaded via importlib — no source changes needed.

Launch

python main.py

Usage

1. Drag the window over the text you want to translate
2. Click ▶ or press Ctrl+L to start
3. Click the language button to choose your target language
4. Translation starts automatically after the window settles (1 second), with change detection every 1.5 seconds

Shortcuts

Shortcut	Action
Ctrl+L	Toggle translation
Escape	Clear and re-detect in 1 second
F5	Force refresh

Notes

• Translation API requires internet (except Ollama local models)
• Default OCR model (Florence-2-base/) must be downloaded separately
• OCR quality is language-dependent — Florence-2 works well for English; use a different engine for other languages
• Close button terminates the process immediately (os._exit(0))