Accurate short-term solar radiation forecasting is essential for the stable operation and dispatch of photovoltaic power generation systems. Advanced encoder–decoder architectures, utilizing satellite remote sensing data, are now primary techniques for this forecasting task. However, these methods encounter significant limitations, particularly as forecast horizons extend. In such scenarios, predictions often exhibit spatial texture degradation and distortions in radiation intensity. This significantly reduces precision and reliability, making it difficult to meet the demands of high-precision applications. To address these limitations, this paper proposes GAN-Solar, a novel quality optimization model for short-term solar radiation forecasting based on Generative Adversarial Networks (GANs). GAN-Solar utilizes an ED-AttUNet model, enhanced with conditional inputs, as its generator. A discriminator, incorporating residual structures, progressive downsampling, and conditional information, is employed to distinguish between real and generated forecasts. This adversarial process, guided by a hybrid loss function and a discriminator treated as a learnable objective function, refines the forecast quality. Experimental results on summer solar radiation data demonstrate that GAN-Solar significantly improves forecast quality. It reduces the Root Mean Square Error by approximately 3.2% and increases the Structural Similarity Index from 0.84 to 0.87 when compared to the baseline ED-AttUNet model. The proposed method effectively mitigates issues of texture degradation and intensity distortion, leading to clearer and more accurate solar radiation predictions.