Generative Artificial Intelligence (AI) has recently advanced medical imaging by realistic image synthesis for data augmentation. In colorectal polyp segmentation, preserving the original image quality during augmentation is essential for reliable model training. We propose a method that integrates the Parallel Reverse Attention Network (PraNet) with the Polyp-Gen synthesis pipeline to generate high-quality synthetic polyp images. Seven publicly available polyp datasets were standardized through a unified normalization pipeline (aspect-ratio-preserving resize to 512 \(\times \) 512 with padding and intensity scaling) and optimized with moderate augmentations, including rotations, flips, scaling, affine transforms, brightness adjustments, and noise injection. We conducted twelve experiments to isolate the effects of augmentation and multi-source training, covering per-dataset, augmented, combined, and combined-augmented configurations. Our method achieved a Dice score of 0.926 and a Jaccard index of 0.873 on the BKAI-IGH NeoPolyp dataset, demonstrating how the generated pictures improve the segmentation performance.