This study introduces SwinPix, a novel network architecture designed to explore the effectiveness of multi-level low-dose (LD) PET inputs as prior knowledge for standard-dose (SD) PET image prediction. By employing SwinPix architecture, we assess the performance of single-input and multi-input models trained with PET data at 4%, 6%, and 10% dose levels. Two models were developed: the first was trained using a single input corresponding to 4%, 6%, and 10% LD PET images, while the second considered a multi-input approach, utilizing three lower-dose inputs to predict the corresponding SD PET images. The performance of the six models was evaluated using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean standardized uptake value (SUV _mean ) bias, SUV _max bias, and root mean square error (RMSE) within the entire head region and malignant lesions. The SwinPix multi-input model outperformed single-input versions across all dose levels. At 4%, PSNR increased by 13%, SSIM improved from 0.97 to 0.99, while RMSE (lesion/head) dropped by 82–86%. Similarly, SUV _mean and SUV _max biases decreased by 78% and 58%, respectively. At 6% and 10% dose levels, SwinPix showed comparable improvements, reducing RMSE by over 40% and SUV biases by up to 53%. Compared to Pix2Pix and Swin Transformer, SwinPix consistently achieved the best reconstruction quality. All improvements were statistically significant ( p < 0.01), supporting the effectiveness of multi-input SwinPix for accurate LD PET imaging. SwinPix offers a hybrid transformer-based solution for LD PET reconstruction. By incorporating multi-level LD PET inputs through a Generative adversarial network (GAN) framework, it enhances image quality and lesion quantification while maintaining computational efficiency, supporting its potential for clinical deployment.