With the increasing market demand for personalized and high-quality food packaging design, traditional design methods gradually expose issues of inefficiency and high costs. To address this, an improved U-Net network model is proposed, which integrates conditional generative adversarial networks (CGANs) and multi-scale feature fusion techniques to efficiently generate packaging design images with consistent styles and rich details. Experimental results demonstrate that the proposed model outperforms traditional models such as convolutional neural networks (CNNs) and Variational Autoencoders (VAEs) in terms of image generation quality and stability. At 150 iterations, the Structural Similarity Index Measure (SSIM) of the proposed model reaches 1.00, and the peak signal-to-noise ratio (PSNR) is improved to 48 dB, exhibiting extremely high reconstruction accuracy. Furthermore, the proposed model exhibits excellent performance in stability tests, with a final stability value close to 0.99, significantly higher than that of the comparison models. In summary, the improved U-Net model provides an efficient solution for food packaging design, enhancing the automation level of the design process while ensuring image generation quality, and opens up new avenues for future intelligent food packaging design.