Feature Request: 2 Step Translation with Online Services and AI Validation

[CoSciBlog]

Version Info

Python version: 3.10.9 (tags/v3.10.9:1dd9be6, Dec 6 2022, 20:01:21) [MSC v.1934 64 bit (AMD64)]
Python executable: C:\100sd\tools\BallonsTranslator-Pro\venv\Scripts\python.exe
Version: 1.7.0
Branch: main
Commit hash: b215d3b
AMD GPU: Yes
Device name: NVIDIA GeForce RTX 3090
Cuda is available: True
Cuda version: 12.8
ZLUDA is available: False
[INFO ] launch:main:275 - QT_API: pyqt6, QT Version: 6.11.0
[INFO ] launch:main:372 - set display language to English
[INFO ] mainwindow:_update:531 - Text detector set to ctd
[INFO ] mainwindow:_update:553 - Inpainter set to lama_large_512px
[INFO ] mainwindow:_update:565 - Translator set to LLM_API_Translator
[INFO ] mainwindow:_update:542 - OCR set to mit48px

Type of Request

New Feature

Area (optional)

None

Description

I've been experimenting with Balloon Translator and Balloon Translator Pro for a few weeks now, and currently I'm using Ollama with Translategemma 27b locally. My translation results aren't the best, and I'm still tweaking the settings. I had an idea: what about a kind of two-step translation process where the texts are first sent to a service like DeepL or Google Translate, and then, as a second step, checked, corrected, or adjusted by LLMs, Chatgpt, or similar services? That might be a nice idea/feature, even if it would be a bit more complex.

Pictures

No response

Additional Information

No response

[thomaswantstobeaskeleton]

May I ask what you are languages you are translating to and from, specific models give different approaches to translation? What you mention has already been partially there, just not in the review sense. I've implemented it and you can select the 2 step translation w review with chain in the translator module selection.

https://github.com/thomaswantstobeaskeleton/BallonsTranslator-Pro#new-2-step-translation-validation-feature-55

[CoSciBlog]

Hi so far i tried a couple of models from 10-30b like gemma3, translategemma, qwen2.5, qwen3.0 with more or less mediocre results, only sometimes a bit better than google translator often worse. tried token size from 4k to 128k, most of the time i use 32-64k because of "better" results

tried ballonstranslator and ballonstranslator pro with rtx 3090 and 5070 ti for translation from japanese to english.
Temperature from 0.1-0.6 and top p from 0.6-1.0

prompt template
`
Translate the following manga/comic dialogue into {to_lang}.

STRICT OUTPUT RULES:

• Output plain text only. Do NOT use Markdown, code blocks, backticks, JSON, YAML, XML, bullet lists, or headings.
• Output EXACTLY one line per input segment.
• Do NOT insert line breaks inside a segment; replace any intended line breaks with a single space.
• Do NOT add blank lines.

SEGMENT RULES (MUST FOLLOW):

• The input contains segment markers like <|1|>, <|2|>, etc.
• Preserve every marker exactly as-is.
• Translate each segment separately.
• Output exactly one translated line for each input marker, in the same order.
• Never merge, omit, reorder, renumber, or alter markers.

TRANSLATION RULES:

• Preserve meaning, tone, and emotional intensity.
• Use natural, grammatically correct {to_lang}, concise for speech bubbles.
• Return the final translation in ALL CAPS.
• Do not add explanations, warnings, labels, or commentary.

Output format:
<|1|>TRANSLATED TEXT
<|2|>TRANSLATED TEXT

Text:
`

chat system template
`
You are a manga/comic dialogue translation engine.

Task:

• Translate the user's text into {to_lang} accurately, faithfully, and naturally.

Output policy (STRICT):

• Plain text only. Do NOT use Markdown, code blocks, backticks, JSON, YAML, XML, bullet lists, or headings.
• Output only translated segments with their original markers. No extra lines, no blank lines.

Consistency policy:

• Keep spelling of names, honorifics, and recurring terms consistent.
• If prior translated context is available, match it exactly.
• If not available, choose one consistent spelling/term and keep it for the entire request.

Speaker/reference handling:

• Use nearby segments to infer speaker and addressee when possible.
• Infer omitted subjects/pronouns for natural {to_lang}; avoid impersonal “IT” when a person is intended.
• Preserve whether the line is a statement, question, command, exclamation, interruption, or vocalization.

Segment rules (STRICT):

• Preserve markers <|N|> exactly.
• Translate only the text after each marker.
• Output exactly one line per marker.
• Never merge, omit, reorder, renumber, or alter markers.
• No line breaks inside a segment; use spaces instead.

Style:

• Concise speech-bubble English.
• ALL CAPS output.
  `

[thomaswantstobeaskeleton]

I recommend checking out https://huggingface.co/cyberagent/CAT-Translate-7b
as it seems to get around the best on translation japanese to english benchmarks.