Libre Manga Translator

Privacy-focused manga translator: local WebGPU, Gemini cloud, or self-hosted LLM

Translate manga directly in your browser - you decide where your data goes. Libre Manga Translator is a privacy-focused, open-source (AGPL-3.0) extension that translates manga pages in-place: detect bubbles, extract text, translate, clean the original text, and typeset the translation back onto the page. THREE MODES, ONE PIPELINE 1. WebGPU Mode (default, fully local, offline after setup) Everything runs on-device: bubble detection + OCR (CPU/WASM) + translation (GPU via WebGPU). No API key, no uploads (into us). Models: Gemma3-1B, Qwen3.5-2B, Qwen3.5-4B. 2. Gemini Mode (cloud, opt-in) Sends the annotated page to Google Gemini with your own API key for single-call OCR + translation. Best for low-GPU machines. Browser -> Google directly, no proxy. 3. API Mode (self-hosted) Connect to Ollama, LM Studio, vLLM, OpenRouter, or any OpenAI-compatible /v1/chat/completions endpoint. Local PaddleOCR extracts text, only text is sent to your server - images never leave your machine. WHAT YOU GET Vision & OCR - 2 on-device detectors: Comic Bubble RT-DETR (default, 11 MB, fast) and Comic Text Detector (94 MB, dense mask seeding) - 3 OCR engines: PaddleOCR multilingual (11 language packs, auto-switched), PP-OCRv6 Manga (21 MB, lightweight Japanese), Manga-OCR (460 MB, vertical-text specialist) - Script gate holds back SFX / art / wrong-language text instead of mistranslating. One-click Translate anyway to override. Clean redraw & typesetting - Fast ladder (default): planar fill -> denoise -> Telea, lightest passing rung wins per region. No ghost boxes. - Quality mode: LaMa neural redraw first, Fast fallback per region for screentones / halftones. - Auto text wrapping + sizing, bundled manga fonts + custom TTF/OTF/WOFF, original/translated toggle, 1-click JPG export. Reader workflow - Hover any image to show Translate pill. After translation it becomes edit / refine boxes / toggle / export hub. - Auto-translate chapters as you scroll, one page at a time. - Drag / resize / add / delete / undo boxes in manga reading order (right-to-left). - Edit OCR + translation in-place without re-running inpainting. - Series title, glossary, and last 5 bubbles carry across pages for consistent names. - Site adapters auto-detect series/chapter/page + AI rule generator for new sites. Control - Model Storage: list / download / update-check / delete all cached weights in one panel. - GPU controls: master WebGPU switch + per-area overrides (LLM / inpaint / OCR) with WASM fallback. - Local-only debug logs (timings, OCR text, inpaint provenance - no image payloads). HOW TO USE 1. Open any manga page 2. Hover image > click Translate pill 3. Adjust boxes if needed (or enable Skip refining / Auto-translate) 4. Confirm > read 5. Edit, Refine, Save JPG, or toggle Original anytime PRIVACY - Zero telemetry. No analytics, no tracking, no uploads in WebGPU Mode. - Gemini / API modes are opt-in only and send data only to the endpoint you configure. - API keys and server settings are stored machine-local, never synced. - Background image fetch is SSRF-gated (blocks loopback / private IPs / metadata) and stays in your browser. - Models download once from Hugging Face, then work offline. Permissions used: activeTab + scripting (render overlay on the manga page you translate), storage + unlimitedStorage (settings + model cache), offscreen (isolated on-device inference), declarativeNetRequest (fix CDN referer / hotlink blocks), host access <all_urls> (translate on any reader site you visit). Open source, works offline after 0.8-4 GB initial model download depending on LLM choice or point it into your local self-hosted model. Please support the official releases of manga you read. For personal use and educational exploration.