Exploiting Relationship for Complex-scene Image Generation

The significant progress on Generative Adversarial Networks (GANs) has facilitated realistic single-object image generation based language input. However, complex-scene (with various interactions among multiple objects) still suffers from messy layouts and object distortions, due to diverse configurations in appearances. Prior methods are mostly object-driven ignore their inter-relations that play a role images. This work explores relationship-aware generation, where objects inter-related as scene graph. With the help of relationships, we propose three major updates framework. First, reasonable spatial inferred by jointly considering semantics relationships objects. Compared standard location regression, show relative scales distances serve more reliable target. Second, since relations between have significantly influenced an object's appearance, design relation-guided generator generate reflecting relationships. Third, novel graph discriminator is proposed guarantee consistency generated input Our method tends synthesize plausible objects, respecting interplay image. Experimental results Visual Genome HICO-DET datasets our outperforms prior arts terms IS FID metrics. Based user study visual inspection, effective generating logical layout appearance for complex-scenes.