This paper introduces a novel framework for machine unlearning in image-to-image (I2I) generative models, addressing the challenge of selectively removing sensitive data samples ("forget samples") while preserving the performance on the remaining data ("retain samples"). The proposed method is computationally efficient and theoretically grounded, ensuring minimal performance degradation on retain samples while effectively removing information from forget samples. The framework is applicable to various I2I generative models, including diffusion models, VQ-GAN, and MAE. The method is validated on two large-scale datasets, ImageNet-1K and Places-365, demonstrating its effectiveness without requiring access to the retain samples. The algorithm is designed to minimize the KL divergence between the distributions of retain and forget samples, with a focus on maximizing the divergence for forget samples. The approach is further enhanced by converting the KL divergence into an L2 loss, making it computationally feasible for diverse I2I generative models. The method is evaluated against several baselines, showing superior performance in terms of image quality and data privacy protection. The results indicate that the proposed approach is robust and effective in unlearning tasks, even when the retain samples are not available. This work represents the first systematic exploration of machine unlearning for I2I generative models, providing a comprehensive framework that addresses both theoretical and practical challenges in the field.This paper introduces a novel framework for machine unlearning in image-to-image (I2I) generative models, addressing the challenge of selectively removing sensitive data samples ("forget samples") while preserving the performance on the remaining data ("retain samples"). The proposed method is computationally efficient and theoretically grounded, ensuring minimal performance degradation on retain samples while effectively removing information from forget samples. The framework is applicable to various I2I generative models, including diffusion models, VQ-GAN, and MAE. The method is validated on two large-scale datasets, ImageNet-1K and Places-365, demonstrating its effectiveness without requiring access to the retain samples. The algorithm is designed to minimize the KL divergence between the distributions of retain and forget samples, with a focus on maximizing the divergence for forget samples. The approach is further enhanced by converting the KL divergence into an L2 loss, making it computationally feasible for diverse I2I generative models. The method is evaluated against several baselines, showing superior performance in terms of image quality and data privacy protection. The results indicate that the proposed approach is robust and effective in unlearning tasks, even when the retain samples are not available. This work represents the first systematic exploration of machine unlearning for I2I generative models, providing a comprehensive framework that addresses both theoretical and practical challenges in the field.