I'm trying to catch backwards "selfie" images before passing them to our VLM text reader and/or face embedding extraction. Since models like Qwen and Florence are trained on flipped data, they are mostly blind to backwards text and prompting them just seems to be fighting against their base training
