Parrot provides a multi-reward reinforcement learning framework that addresses the issues of reward over-optimization and degradation in text-to-image generation. The framework shows significant improvements across various quality metrics, including aesthetics, image sentiment, and human preferences, using a joint optimization method. By introducing an innovative prompt-centric guidance, it effectively tackles the over-optimization problem and generates images faithful to the original prompts. While Parrot excels in image quality, its reliance on existing metrics raises ethical concerns. Parrot's adaptability enhances the quantification of image quality.