Using the flutterfire sdk for gemini ai in a flutter application.
With a regular call to gemini ai model = FirebaseAI.googleAI().generativeModel(model: 'gemini-2.5-flash'); and with a screenshot of my desktop and recorded audio instructions (“list the names of the chat rooms visible on my screen” or “what is the name of the app that is being developed in visual studio code on my screen”, gemini provides accurate answers.
However, when using the live api liveModel => FirebaseAI.googleAI().liveGenerativeModel(model: 'gemini-live-2.5-flash-preview'), with the same instructions and prompts, the responses are completely incorrect. Sometimes they seem hallucinated with no basis in the image, sometimes there are correct portions intermingled with others. The live api has my screenshot and can identify which apps are in which portions of my screen, although it has incorrectly identified which IDE and messaging apps I am using at times.
I am sending the same screenshot data to each model and preview that data to know that it is a high quality image.
It seems that somewhere along the live api processing, the image gets significantly distorted and text recognition becomes near impossible.