Farshad Azade ; Musatian Sabrina ; Dhamo Helisa ; Navab Nassir

Abstract

Generation of images from scene graphs is a promising direction towardsexplicit scene generation and manipulation. However, the images generated fromthe scene graphs lack quality, which in part comes due to high difficulty anddiversity in the data. We propose MIGS (Meta Image Generation from SceneGraphs), a meta-learning based approach for few-shot image generation fromgraphs that enables adapting the model to different scenes and increases theimage quality by training on diverse sets of tasks. By sampling the data in atask-driven fashion, we train the generator using meta-learning on differentsets of tasks that are categorized based on the scene attributes. Our resultsshow that using this meta-learning approach for the generation of images fromscene graphs achieves state-of-the-art performance in terms of image qualityand capturing the semantic relationships in the scene. Project Website:https://migs2021.github.io/