Classify and Generate: Using Classification Latent Space Representations for Image Generations

Utilization of classification latent space information for downstream reconstruction and generation is an intriguing a relatively unexplored area. In general, discriminative representations are rich in class specific features but too sparse reconstruction, whereas, autoencoders the dense has limited indistinguishable features, making it less suitable classification. this work, we propose modelling framework that employs manipulated supervised to reconstruct generate new samples belonging given class. Unlike generative approaches such as GANs VAEs aim model data manifold distribution, Representation based Generations (ReGene) directly represents space. Such representations, under certain constraints, allow reconstructions controlled generations using appropriate decoder without enforcing any prior distribution. Theoretically, class, show these when smartly convex combinations retain same label. Furthermore, they also lead novel visually realistic images. Extensive experiments on datasets varying resolutions demonstrate ReGene higher accuracy than existing conditional models while being competitive terms FID.