An Amazing Hidden Gem! (with feedback & a feature suggestion)

[ginzbitcoin]

Hello!

I just wanted to say that your application is absolutely fantastic. It is much lighter and far less complex than other tools I have used in the past. I'm so happy to have found a hidden gem like this! The responsiveness is great and the translation quality is excellent.

I did want to provide some small feedback on the OCR performance. I've noticed that during real-time translation (DeepL), it sometimes captures random characters or code along with the actual text (for example, things like : AX or A1()==21xba). This can make the final translation output a little unclean. I understand that this is a common challenge for all OCR technology, but any future improvements in this area would make the tool even more perfect.

On that note, I also had a feature suggestion regarding the capture boxes. I was hoping it might be possible to have multiple OCR and Translate boxes running on the screen at the same time.

For example, in many complex games, it would be incredibly helpful to have:

1. One OCR box capturing the character dialogue at the bottom of the screen.
2. A second OCR box simultaneously capturing the quest log on the right side.
3. And maybe even a third for item stats that pop up in another corner.

Being able to set up and run multiple, simultaneous translation zones like this, instead of just a single pair of boxes, would be a true game-changer.

Thank you again for creating such a wonderful and helpful tool. Keep up the great work!

[tomkam1702]

Thank you for your kind words.

Here are a couple of tips for dealing with unwanted characters in your output:

This is a common problem when using Tesseract for OCR. If you use a Gemini or GPT model instead, you should see far fewer of these strange characters.

Your second line of defence is an LLM-based translation model, like Gemini or GPT. If some unwanted characters are still captured, even when using a modern OCR tool like Gemini 2.0 Flash, an LLM translation model is often smart enough to recognise that these characters don't belong and will leave them out of the translation. This is something that tools like DeepL or Google Translate might not be able to do.

In short, for the best translation quality, use LLM models for both your OCR and translation needs.

[ginzbitcoin]

This is super helpful, thank you! And I just checked... turns out I was still on an older version, so I hadn't seen the new OpenAI options yet. I'm excited to update and try it out!