Generated Contents Enrichment: Explicitly Enriching Image Generation Through Scene Graphs

The paper introduces Generated Contents Enrichment (GCE), a conditional image-generation task that makes scene enrichment explicit before rendering. A sparse description is converted into a scene graph, graph convolutional networks predict additional objects and inter-object relations, and the enriched graph is passed to a downstream image-generation pipeline. The jointly trained adversarial framework is evaluated on Visual Genome using proxy scene-graph enrichment metrics, image-quality comparisons, qualitative examples, and user studies. The central goal is to produce visually plausible and structurally coherent images whose content is semantically richer than the original sparse description.