The article presents a new filter pruning method tailored for deep models that generate images. Unlike traditional approaches that rely on pixel intensity in intermediate activations, the proposed method considers the spatial position of pixels in the image and employs binary masks to distinguish essential areas (objects) from the background. This technique allows for assessing filter importance by focusing on significant regions within the feature maps. The article compares several pruning criteria, demonstrating that our approach achieves high compression rates while maintaining a low mean squared error on reconstructed images.