Japanese Manga Machine Translation Experience (5) - ImageTrans

About 1581 wordsAbout 5 min

Machine-Translation Translation Manga AI

2025-07-31

This article is translated by AI.

ImageTrans is a paid translator, don't ask me if there is a cracked version.

Archive Directory

Japanese Manga Machine Translation Experience (0) - Table of Contents

Introduction and Purchase Address

ImageTrans

After successful purchase, an email will be sent to your mailbox. Pay attention to check if it is judged as spam.

Usage Suggestions

Startup

It is recommended to double-click the ImageTrans.bat file to start the translator, so that you can see the running status of the translator, at least you can know if the software is stuck...

Plugins

"Tools" -> "Plugin Manager" in the top bar, install the following three according to the installation instructions.

Japanese Manga Machine Translation Experience (5) - ImageTrans-1753961287746

Bubble Detection Model

Refer to balloon-dataset, add as needed.

It is recommended to add the ysgyolo_S150best bubble detection model. For a brief introduction and effect, see Japanese Manga Machine Translation Experience (3) - BallonsTranslator.

Download the compressed package in the link below, and then unzip it to the model folder.

https://wwke.lanzoue.com/iEelI329p5wh

Check if it is successfully installed in "Bubble Detection" in "Preferences".

Japanese Manga Machine Translation Experience (5) - ImageTrans-1753961860789

chatGPT prompt

prompt

Your primary and mandatory task is to translate the following text to {langcode}.

**Rules:**
1.  **Translate, Don't Echo:** You MUST provide a translation. Do not just output the original text.
2.  **Style:** The translation must be colloquial, elegant, and fluent, avoiding a machine-translated feel.
3.  **Integrity:** Only translate the text content; do not add any interpretation or explanation.
4.  **Formatting:** Preserve all original special characters exactly as they are.
5.  **Exception Clause:** Only in the rare case that the source text is a truly random and meaningless string (e.g., "ajg@#kjl_"), may you return it as is. Do not mistake loanwords, names, or technical terms for gibberish.

**Text to translate:**
{source}

batch_prompt

Your task is to translate all translatable string values within the following JSON object into {langcode}. You must return a valid JSON object with the identical key structure.

**Translation Rules for JSON values:**
1.  **Translate, Don't Echo:** You MUST translate every value. Do not just echo the original value.
2.  **Style:** The translation must be colloquial, elegant, and fluent, avoiding a machine-translated feel.
3.  **Integrity:** Only translate the text content; do not add any interpretation or explanation.
4.  **Formatting:** Preserve all original special characters within each value.
5.  **Structure:** Do not mix or merge values from different keys.
6.  **Exception Clause:** Only in the rare case that a value is a truly random and meaningless string (e.g., "ajg@#kjl_"), may you leave it unchanged. Do not mistake loanwords, names, or technical terms for gibberish.

**JSON to translate:**
{source}

prompt_with_term

Your primary and mandatory task is to translate the following text to {langcode}, strictly using the provided terms.

**Mandatory Terms (JSON):**
{term}

**Rules:**
1.  **Translate, Don't Echo:** You MUST provide a translation. Do not just output the original text.
2.  **Term Usage:** You must use the provided terms where applicable.
3.  **Style:** The translation must be colloquial, elegant, and fluent, avoiding a machine-translated feel.
4.  **Integrity:** Only translate the text content; do not add any interpretation or explanation.
5.  **Formatting:** Preserve all original special characters exactly as they are.
6.  **Exception Clause:** Only in the rare case that the source text is a truly random and meaningless string (e.g., "ajg@#kjl_"), may you return it as is. Do not mistake loanwords, names, or technical terms for gibberish.

**Text to translate:**
{source}

batch_prompt_with_term

Your task is to translate all translatable string values within the following JSON object into {langcode}, strictly using the provided terms. You must return a valid JSON object with the identical key structure.

**Mandatory Terms (JSON):**
{term}

**Translation Rules for JSON values:**
1.  **Translate, Don't Echo:** You MUST translate every value. Do not just echo the original value.
2.  **Term Usage:** You must use the provided terms where applicable in the translation.
3.  **Style:** The translation must be colloquial, elegant, and fluent, avoiding a machine-translated feel.
4.  **Integrity:** Only translate the text content; do not add any interpretation or explanation.
5.  **Formatting:** Preserve all original special characters within each value.
6.  **Structure:** Do not mix or merge values from different keys.
7.  **Exception Clause:** Only in the rare case that a value is a truly random and meaningless string (e.g., "ajg@#kjl_"), may you leave it unchanged. Do not mistake loanwords, names, or technical terms for gibberish.

**JSON to translate:**
{source}

Custom Workflow

The usage method I figured out myself is "Custom Workflow".

Japanese Manga Machine Translation Experience (5) - ImageTrans-1753961907214

The configuration file of "Custom Workflow" is the custom_workflow.json file in the root directory of the manga to be translated. If not, you can create a new one.

Workflow based on manga-image-translator. Good effect, slow speed.

{
    "ocrInterval": 10,
    "engine": "chatGPT",
    "panel_detection_method": 0,
    "name": "MIT Workflow",
    "mergeOperation": "2",
    "detectionMethod": "0",
    "scale": 0.5,
    "interval": 200,
    "type": "MT",
    "auto_remove_low_confidence_areas": true,
    "play_sound": false,
    "flow": [
        "Text Detection (0,db_detector+rotationdetection (mangaTranslator),4)",
        "Generate masks and text-removed images for all images",
        "Generate regions from masks",
        "Merge regions",
        "Shrink regions",
        "Sort",
        "Detect text direction and set corresponding style",
        "Text Recognition (manga-ocr,0)",
        "Resize according to mask",
        "Remove non-original text areas from all images",
        "Translate",
        "Translate",
        "Translate",
        "Delete mask images",
        "Generate finished images"
    ]
},

Workflow based on ysgyolo_S 150 best and mangaocr. Poor effect, fast speed.

{
    "ocrInterval": 10,
    "engine": "chatGPT",
    "panel_detection_method": 0,
    "name": "ysgyolo+mangaocr",
    "mergeOperation": "2",
    "detectionMethod": "3",
    "scale": 0.5,
    "interval": 200,
    "type": "MT",
    "auto_remove_low_confidence_areas": true,
    "play_sound": false,
    "flow": [
        "Text Detection",
        "Generate masks and text-removed images for all images",
        "Generate regions from masks",
        "Merge regions",
        "Expand regions",
        "Shrink regions",
        "Sort",
        "Text Recognition",
        "Detect text direction and set corresponding style",
        "Resize according to mask",
        "Translate",
        "Translate",
        "Translate",
        "Translate",
        "Delete mask images",
        "Generate finished images"
    ]
},

Workflow combining manga-image-translator and ysgyolo_S 150 best detection function, better effect, slower speed.

{
    "ocrInterval": 10,
    "engine": "chatGPT",
    "panel_detection_method": 0,
    "name": "MY Workflow",
    "mergeOperation": "2",
    "detectionMethod": "0",
    "scale": 0.5,
    "interval": 200,
    "type": "MT",
    "auto_remove_low_confidence_areas": true,
    "play_sound": false,
    "flow": [
        "Use bubble detection to remove excess areas and supplement missing areas",
        "Text Detection",
        "Generate masks and text-removed images for all images",
        "Generate regions from masks",
        "Merge regions",
        "Expand regions",
        "Shrink regions",
        "Sort",
        "Panel Detection",
        "Panel Sorting",
        "Clear Original Text",
        "Add text to existing text boxes with OCR",
        "Detect text direction and set corresponding style",
        "Resize according to mask",
        "Remove non-original text areas from all images",
        "Translate",
        "Translate",
        "Translate",
        "Delete mask images",
        "Generate finished images"
    ]
},

For model comparison, please refer to the previous articles in this series.

Japanese Manga Machine Translation Experience (2) - Saber-Translator

Japanese Manga Machine Translation Experience (3) - BallonsTranslator

Japanese Manga Machine Translation Experience (4) - manga-image-translator

Usage Tips

Create Project from Template

If you find it troublesome to reconfigure every time, you can create a project from a template. The template file is in the templates folder and can be manually edited.

Japanese Manga Machine Translation Experience (5) - ImageTrans-1753963979369

Usage Experience

ImageTrans basically cures all the pain points of the translators in the previous three articles, and is the best manga translator I know so far.

If I have to pick a few shortcomings, it is paid, difficult to use, and easy to crash or freeze (this is also the core reason why I gave up ImageTrans and turned to manga-image-translator).

First of all, the plugin sometimes disconnects suddenly, that is, the terminal does not terminate, but the client shows no connection. Or it's not a sudden disconnection, it just can't connect from the beginning.

Then the memory and CPU usage are too high. Although it is recommended to install LamaInpaint, I actually don't use this to repair images, because if I use this to repair three or four hundred pictures, the progress bar will get stuck, the computer will overheat, and then the program will freeze and exit.

There are also cases where other steps freeze, such as panel detection, etc. In short, it is not smooth and very worrying. After all, manga-image-translator has no problem translating one or two thousand pictures continuously.

Finally, I chose manga-image-translator...