فهرست نشریات علمی دسترسی آزاد ایران

Generative Adversarial Networks (GANs) have become influential tools in unsupervised image generation, significantly impacting fields like computer vision and the creative arts. However, challenges such as training instability and mode collapse often hinder their performance and the quality of generated images. This study introduces a hybrid GAN architecture that combines techniques from AdaptiveMix, which enhances model stability, with insights from EIGGAN, known for its innovative image generation methods. The primary goal is to improve both training stability and the visual quality of generated images. The generator incorporates differentiable data augmentation (DiffAugment) and Exponential Moving Average (EMA) updates. DiffAugment introduces dynamic transformations to training data, enhancing diversity and robustness, while EMA updates stabilize training by smoothing parameter changes, resulting in more consistent outputs. The discriminator is regularized using the R1 penalty, improving its ability to distinguish between real and generated images, and benefits from feature space shrinkage through AdaptiveMix to maintain compact feature representations. Trained on the CIFAR-10 dataset for 1000 epochs, the model achieved a peak Inception Score (IS) of 6.80 ± 0.22, indicating significant improvements in generative quality and diversity, along with a best Fréchet Inception Distance (FID) score of 20.90, reflecting high realism in generated samples. The discriminator's accuracy remained stable between 50% and 60%, suggesting a balanced adversarial relationship. These findings demonstrate the effectiveness of the hybrid model. They also open new directions for improving stability and image quality in GAN training.