Controllable Image Synthesis via SegVAE

MeronymNet

“…In some cases, we modified existing scene generation approaches having layouts as the starting point [1,22,32] to generate objects. We also included modified variants of two existing part-based object generative models (3-D objects [34], faces [7]). To evaluate individual components from MeronymNet, we also designed hybrid baselines with MeronymNet components (BoxGCN-VAE, LabelMapVAE, La-bel2obj) included.…”

“…2-D object generative models: Within the 2-D realm, approaches are designed for a specific object type (e.g. faces [2,7], birds [40,42], flowers [28]) and usually involve text or pixel-level conditioning [15]. In some cases, part attributes are used to generate these specific object types [12,36].…”

“…The first stage produces object bounding box layouts while the second stage component [32,45] translates these layouts to scenes. In some cases, the first stage produces semantic maps [1,7] which are translated to scenes using pixel translation models [15,29]. We too employ the approach of intermediate structure generations 1 Meronym is a linguistic term for expressing part-to-whole relationships.…”

“…Many recent successes for generative approaches have been associated with generation of realistic scenery using top-down guidance from text [14] and semantic scene attributes [1,29,30,32,45]. Alongside, generative models for individual objects have also found a degree of success [2,4,7,12,40,42]. Despite these successes, the degree of controllability in these approaches is somewhat limited compared to the scene generation counterparts.…”

“…The bulk of object generation approaches specialize for a single category of objects with weakly aligned part configurations (e.g. faces [2,7,12]) and for objects with associated text description (e.g. birds [40,42]).…”