Well I am comfortable enough to release this as it is the prompts need a lot of work but are "usable" (would love pull requests for new ones and improvements). I still plan to work on this and make improvements.


Here is #PixelPolygot 0.5.0 my solution for roughing your way through a Japanese video game when you only speak English.

https://codeberg.org/Melon-Bread/Pixel-Polygot

(with a GH mirror in case codeberg blows up 🙄 https://github.com/Melon-Bread/Pixel-Polygot )

Will end up releasing all the binary builds I am able to make on next release once iron out a few things can streamline the build process with so many targets for my self.

CC: @sun

@lain Yeah like I mentioned before it is a ongoing thing and most LLMs are just not cut out for it since it is extracting Japanese text in most cases. It was not till I started playing with the Qwen2.5-VL models that I thought that this could actually work to some degree.

@sun @lain The more I have used and tested things
I have come to same conclusion or at least get more specific with the type of game you are working with in the prompt. Like fleshing out more a rpg prompt vs a visual novel and so on. You could come at it quite big context wise out the gate since each image is a new conversation.

I like the idea of having an ongoing context for chunks of the game in scenes to make it flow better or as previous mentioned more consistent though not sure what is the best way to go about it.