<p>Image restoration is a core research direction in the field of machine vision. Traditional restoration methods suffer from problems such as blurry details, unnatural textures, and difficulty in handling complex structures. Existing techniques based on generative adversarial networks face challenges such as unstable training and inaccurate restoration of details in repairing large areas of missing images. This study proposes an image inpainting technique based on a generative adversarial network model. The network generator incorporates adaptive feature blocks and a Transformer to optimize feature extraction and long-range dependencies. An image inpainting algorithm with an improved denoising diffusion probability is proposed, enhancing the network’s ability to inpaint images with large areas of missing content. The recall rate of the model on the CelebA dataset reaches 94.2%, with an accuracy rate of 97.2%; The structural similarity index of 30% and 50% missing images is superior to similar techniques such as Style GAN and RestoreFormer++; In the restoration of the Places2 dataset with a missing rate of 50–60%, the peak signal-to-noise ratio reaches 23.4dB and the average absolute error is 0.038, which can accurately restore the details and textures of the missing areas. In 50% to 60% of missing image restoration, the research model can more accurately restore missing details and textures in the image, and its overall performance is better than similar techniques. This technology is suitable for image restoration needs in different scenarios such as low to medium area and large area defects. It effectively solves the application shortcomings of traditional methods and existing GAN models. This research provides core support for the development and implementation of high-performance image restoration technology, and promotes the development of artificial intelligence in the field of image processing.</p>