Open Source Browser Extension

Manga Translation,
Powered by Local AI

Speech bubbles detected, text extracted, and translated entirely on your device.
WebGPU acceleration. No uploads. No accounts. No subscriptions.

ComicTL
Demo

What it does

Built for privacy. Built for speed.

No manga pages go anywhere. Detection, OCR, and translation run where you are: in the browser, on your hardware.

100% Local Inference

Your manga never leaves the browser. Detection runs through YOLO-Nano via ONNX Runtime Web. Translation runs on WebLLM, accelerated by WebGPU. No server, no API key required, no subscription.

WebGPU WebLLM On-device

Lightning Fast Vision

A custom YOLO-Nano model trained on 5,595 manga pages finds speech bubbles in milliseconds. Trained on Manga109-s and MangaDex data. PaddleOCR pulls the text out on-device, no round-trips needed.

93%

Precision

94.7%

mAP@50

2.4M

Params

Cloud Option Available

Prefer cloud-quality output? Bring your own Gemini API key and switch to Cloud Mode. The annotated page image is sent directly to Gemini, which handles OCR and translation in one step. Your API key never touches our servers — the request goes straight from your browser to Google.

Your key. Direct to Google. No proxy.

Smart Series Context

Set a title, plot summary, and custom dictionary per series. Character names and terminology stay consistent across chapters. The last five translations are used as context for subsequent pages.

Custom dictionary Chapter history Per-series settings

Usage

Four steps. No setup.

Open a manga page, click the icon, and you are reading in your language within seconds.

01

Open any manga page

Navigate to any manga page in Chrome or Firefox. Right click and select "Translate Image" with ComicTL icon in the context menu.

Step 1
02

Bubbles detected automatically

YOLO-Nano runs locally and draws numbered bounding boxes around every speech bubble. Detection happens in under a second on most machines.

Step 2
03

Adjust the boxes if needed

Drag, resize, add, or delete boxes before confirming. Undo and redo are fully supported. When the boxes look right, click Confirm.

Step 3
04

Translation painted in-place

In Local Mode, PaddleOCR extracts the text and a local Llama model translates it on-device via WebGPU. In Cloud Mode, the annotated image goes to Gemini which handles both. Either way, the result is painted directly onto the original page. Toggle between original and translated at any time.

Step 4

Under the hood

A real inference pipeline in the browser

No toy wrappers. ONNX Runtime Web runs YOLO directly. PaddleOCR extracts text on-device. WebLLM loads actual language model weights and runs them on WebGPU.

Detection Layer

YOLO26-Nano / Small

ONNX Runtime Web

PaddleOCR

On-device text extraction

Offscreen Document

Isolated inference thread

Translation Layer

WebLLM

Qwen3 4B (Balanced) · Qwen3 8B (High Quality)

Gemini API

Optional. Your key only.

Series Context

5-chapter rolling history

Extension Layer

WXT Framework

Chrome + Firefox builds

Svelte 5 Runes

Fine-grained reactivity

Supabase

Opt-in bbox telemetry

Read manga your way. No cloud tax.

Free and open source. MIT licensed. No account. No API key required to get started.