This paper presents an enhanced Pix2Pix Generative Adversarial Network (GAN) designed for robust and efficient object removal in street-view imagery. To address the challenges inherent in street-view images such as geometric distortions and high computational complexity, we propose key modifications to the traditional Pix2Pix framework. First, we employ a viewport projection module to isolate and rescale target objects, effectively reducing background interference and image distortion. Next, we introduce an adaptive loss balancing strategy that dynamically calibrates the weights of adversarial, perceptual, style, total variation, and pixel-wise losses during training. This strategy leverages an Exponential Moving Average (EMA) to enhance training stability and improve output quality. Furthermore, we incorporate a clipped adversarial loss mechanism to ensure consistent gradient feedback by preventing the adversarial loss from diminishing excessively. Extensive experiments conducted two distinct urban street-view datasets, including Bride and City dataset, demonstrate that our method significantly outperforms baseline models, including Pix2Pix and a fixed-weight multi-loss Pix2Pix variant, where the best results on Bridge with MSE 159.33, PSNR 26.19, and SSIM 0.91, and on City with MSE 165.44, PSNR 28.98, and SSIM 0.84, demonstrating its superior performance in urban object removal tasks.