Mixgan: çeşitli ve düzenlenebilir içboyama için çift yollu stylegan füzyonu

While generative models have advanced image inpainting, achieving diverse, editable results that blend seamlessly remains a key challenge, particularly for GANs, which struggle to balance reconstruction fidelity with editability. This thesis introduces a novel framework to address these gaps, unifying diversity, fidelity, and user control within a single GAN-based system. First, a new architecture leverages StyleGAN's feature space by employing parallel generation pathways—one from the encoded image and another from a random latent sample. Through novel mixing networks with dual feedback, these pathways are dynamically aligned in the feature space, rather than the latent space, enabling diverse, high-quality inpaintings that maintain structural integrity. Next, the framework perfects the fidelity-editability balance using masked skip connections with an information bottleneck. This forces the model to rely on editable low-rate codes for missing regions while using high-rate features to flawlessly reconstruct visible parts. The result is seamless boundary transitions and preserved control for attribute-specific edits, overcoming common artifacts in prior work. Together, these contributions provide a robust solution that advances GAN-based inpainting. The method generates diverse, high-fidelity, and editable content by operating in the feature space and managing information flow. Extensive experiments demonstrate that our framework surpasses existing approaches in both quality and diversity, offering a fast, controllable system for image restoration and editing.