PRISM: Progressive Restoration for Scene Graph-based Image Manipulation

Scene graphs have emerged as accurate descriptive priors for image generation and manipulation tasks, however, their complexity and diversity of the shapes and relations of objects in data make it challenging to incorporate them into the models and generate high-quality results. To address these challenges, we propose PRISM, a novel progressive multi-head image manipulation approach to improve the accuracy and quality of the manipulated regions in the scene. Our image manipulation framework is trained using an end-to-end denoising masked reconstruction proxy task, where the masked regions are progressively unmasked from the outer regions to the inner part. We take advantage of the outer part of the masked area as they have a direct correlation with the context of the scene. Moreover, our multi-head architecture simultaneously generates detailed object-specific regions in addition to the entire image to produce higher-quality images. Our model outperforms the state-of-the-art methods in the semantic image manipulation task on the CLEVR and Visual Genome datasets. Our results demonstrate the potential of our approach for enhancing the quality and precision of scene graph-based image manipulation.