Skip to content

Quick Start Guide

Neko Translator is an AI-powered assistant for translating comics and manga.

Whether you are a reader translating foreign-language comics or a professional translator, you can use it as a platform for your translation work.

This guide walks you through installation, configuration, and connecting to online model APIs.

Neko Translator itself is free, but the third-party model APIs it connects to may charge you. In our tests, typical translation jobs do not cost much:

  • With gemma4-31b-it from OpenRouter as the vision model (reasoning disabled), recognizing around 1,000 text boxes/images costs about US$0.089.
  • With deepseek-v4-flash from DeepSeek’s official API (reasoning enabled), translating 50 comic pages costs about US$0.074.
  • If you have a powerful GPU, connecting a local model is also a good way to avoid these costs.

Note: Run a small test translation first and keep an eye on your token usage so you don’t run up unexpected charges. If your provider supports it, we also recommend setting a spending limit on the API key.

Visit our website in a browser to download Neko Translator.

This guide recommends and uses google/gemma-4-31b-it as the vision model and deepseek/deepseek-v4-flash with reasoning enabled as the translation model.

google/gemma-4-31b-it is a low-cost choice that handles Extract Text well, but its image summaries are only mediocre. If you rely on Extract Image Summary for translation context, consider a more capable vision model; see Service Configuration: Generic AI.

  1. Visit OpenRouter and sign in or create an account.
  2. If you plan to use paid models, add funds on the Credits page.
  3. Open the Keys page and select Create Key.
  4. Give the key a recognizable name, create it, then copy the key and store it securely. You will only need to paste it into Neko Translator once.

Return to Neko Translator and open Service Profiles.

Service Profiles in landscape layout

First, configure OpenRouter API:

  1. Under General AI API, select + Add and choose OpenRouter.
  2. Paste the API key beginning with sk that you copied earlier.
  3. Keep the preconfigured model preset unchanged.
  4. Save the profile.
  1. Open Service Profiles.
  2. Under Text Block Detection, find Local Detect Text Block Model and select Download.
  3. Under Text Extraction, find Local Text Extractor and select Download. Start with the Medium model. If it runs slowly, switch to a smaller version.

You can also download these models from other sources, such as the Neko Translator homepage. Select Import zip to import a model archive.

Download options for the local text block detector and text extractor

  1. After setup, the local text block detector and text extractor should appear as shown below.

Configured local text block detector and text extractor

After configuration, test that each stage works as expected:

  1. Add a comic, then upload one or more comic pages.
  1. Click Detect Text Blocks in Image. When the text boxes appear, the text block detection service is configured correctly.
  1. Click Extract Text from Text Blocks. When text appears in the text blocks, the text extraction service is configured correctly.

    Note: The Local Text Extractor supports Korean starting with version v1.4.0. If you are using an earlier version, upgrade to recognize Korean.

  1. Click Extract Image Summary. A check mark on the icon confirms that the model for Extract Image Summary is configured correctly.
  1. Click Translate All Text Blocks. If translations are generated, the translation model is configured correctly.

Comic source: Say Hello to Black Jack by Shuho Sato.

Once you have confirmed that the configuration works, choose the images and automation steps you want to run.

  1. Click the ✨ icon to open Batch Processing. By default, all images in the current chapter are selected; choose the automation steps you want to run.
  2. Select the steps you need, then click Start and wait for processing to finish. Image parsing is optional; use it when image descriptions are useful as translation context.

Tip: The current setup uses the Local Text Block Detector. It finds text regions but does not extract their text. Run Extract Text from Text Blocks when you need to populate source text. Image parsing is optional.

Models such as PaddleOCR VL 1.6 can also return recognized text during detection. By default, the separate text-extraction step fills only text blocks whose source text is empty; choose the All images scope only when you intend to re-run extraction on existing text.

The comic translation workflow consists of four stages:

  1. Detect Text Block finds text regions and creates text blocks. Some models also Extract Text during detection.
  2. Extract Text reads text from the image and saves it as the source text in each text block. This requires a vision-capable multimodal model or a dedicated OCR model.
  3. Extract Image Summary (optional) describes the image contents in text to provide context for translation. This requires a vision-capable multimodal model.
  4. Translate uses that context to translate the source text in detected text blocks. This requires a large language model.

These stages correspond to automation buttons in the comic editor. Available actions depend on the capabilities of the models configured in Service Profiles:

Automation capabilities

Generic AI connects to large language model APIs:

  • Generic AI models can Translate.
  • Models with vision capabilities can also Extract Text and Extract Image Summary.

Detect Text Block requires specially supported models. The currently supported options are:

  • Local text block detector: download the model to find text regions on comic pages. It works offline and is free to use. A separate model is needed to extract text. This is the recommended option.
  • PaddleOCR VL 1.6: requires an API and can also extract text during detection.
  • GLM-OCR: requires an API and can also extract text during detection.

Extract Text also has a specially integrated option:

  • Local Text Extractor: download the model to extract text directly on the device. It works offline and is free to use.

If you run into configuration problems or have suggestions, join the community discussion on Discord.

Discord