This article is translated by AI.

For manga translators, I introduce Saber-Translator first because it is the easiest to deploy, has detailed tutorials, is easy to get started with, and I have also accumulated some experience in the QQ group.

Archive Directory

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

Saber-Translator Project Address

MashiroSaber 03/Saber-Translator

The tutorial is very detailed. If you have any questions, check the tutorial and issues first.

Resource Transfer

Files shared via network disk: Saber-Translator Link: https://pan.baidu.com/s/1RnWoKy2tDboWUgsJ5gfsjg?pwd=kvfx Extraction code: kvfx

Installation Suggestions

Use a foreign network environment throughout the process, otherwise some python packages may not be downloaded.

Usage Suggestions

OCR

For OCR, mangaocr is recommended for Japanese manga. AI visual OCR is not necessary for daily use because it is slow. And the effect may not be as good as mangaocr, it is only suitable for solving difficult problems. For others, you can try paddleocr.

It should be noted that mangaocr is only suitable for Japanese manga. For Japanese manga, the effect of mangaocr is far better than paddleocr. If it is a relatively obscure language, you can try AI visual OCR.

In addition, the CPU version of Baidu OCR seems to be broken and cannot be used, but the GPU version can still be used normally.

Translation

Deepseek V 3 is not highly recommended, because I tried several sets of prompt words, and I always encountered the problem of annotations appearing in the translation. Of course, if you find a good set of prompt words, you can also use Deepseek v 3.

Note: According to group friend 业余的教授, the problem of Deepseek V 3 annotations can be solved by adding the following paragraph to the prompt words.

Please note that you can only output the translation of the original text and not include any irrelevant content such as annotations or thinking processes.

Translation models recommend gemini-2.5-flash-lite-preview-06-17 and gemini-2.5-flash. These two models have free quotas in google AI studio. And the effect is good.

High Quality Translation

It seems that only gemini-2.5-flash and gemini-2.5-pro, as well as doubao 1.6 seed can be used, and the effects of others are not ideal.

Note: According to the suggestion of group friend 浅浅晚星, gpt-4.1-mini is also okay.

Usage Experience

In addition to being easy to deploy and having detailed tutorials, Saber-Translator also features "high-quality translation".

But the disadvantage is that it cannot translate in large batches. Many group friends in the QQ group reported that if too many pictures are translated at once (but still less than 100), they cannot be packaged and downloaded. Of course, some group friends have successfully downloaded 200 pictures at once.

In addition, another problem is that you cannot freely choose the model for the "detection" step (What does "detection" mean? See Japanese Manga Machine Translation Experience (1) - Overview), so there will be cases where text outside the bubbles, and text areas such as rotated bubbles cannot be detected (especially long strip Korean manga). In this case, you can only manually label it, which is very troublesome. Especially like the one below where a page is full of background text.

In addition, basic settings, such as apikey, seem unable to be saved. You need to fill them in again every time you open the software, which is also quite troublesome.

Another point is that if "high-quality translation" is not enabled, the prompt for LLM translation only has the text of one text area. Saying this may not reflect the problem, but comparing it with manga-image-translator and BallonsTranslator will make it clear. The latter sends the text areas of the same picture to the LLM together as a prompt after sorting (panel detection and other steps). Therefore, Saber-Translator is equivalent to missing context and is just single-sentence translation.

FAQ

How to Ask Questions

When asking questions, please be sure to include the following information (screenshots, do not take photos of the screen):

0. Detailed description of the problem, including key operation steps and results
1. Downloaded version (How many files were downloaded? Where were they downloaded from? CPU or GPU?)
2. Complete installation path (absolute path)
3. Network environment, what node
4. Terminal (black window, never close it!) yellow or red warning situation
5. All configuration information, such as large model supplier, OCR model, etc.
6. GUI interface (interface opened by browser) error reporting situation
7. Computer configuration, such as graphics card model, cuda version
8. Error pictures

Which one to choose, CPU version or GPU version

Users with N cards with more than 8 G of video memory are recommended to choose the GPU version, and others are recommended to choose the CPU version. The CPU version is universal.

High Quality Translation Model Selection

Do not use models that do not have multimodal capabilities such as Deepseek V 3 or Deepseek R 1!!!

Otherwise, the following error may occur.

Symbols like ♡ cannot be displayed in the translation

Change to a font that supports these symbols.

What to do if stuck on this page

Disconnect from the network and restart the software. After loading successfully, reconnect to the network.

Failed to parse JSON result returned by AI

Common in high-quality translation using Gemini 2.5 pro, just turn off Force JSON Output.

High-quality translation using doubao 1.6 seed reports an error

Solution: Convert the image format to be translated to png format.

How to manually label

From group friend 江夏

Contributors: Ming Wei

Previous pageJapanese Manga Machine Translation Experience (3) - BallonsTranslator

Next pageJapanese Manga Machine Translation Experience (1) - Overview