LR-GAN: Layered Recursive Generative Adversarial Networks for Image Generation

TL;DR

LR-GAN introduces a layered recursive approach to image generation, enabling more natural and recognizable images by separately modeling background and foregrounds and stitching them contextually.

Contribution

It proposes a novel layered recursive GAN architecture that generates scene components separately and combines them, improving image realism and object recognizability.

Findings

LR-GAN produces more natural images than DCGAN.

The model generates objects with clearer shape and pose.

Unsupervised end-to-end training is effective for complex scene generation.

Abstract

We present LR-GAN: an adversarial image generation model which takes scene structure and context into account. Unlike previous generative adversarial networks (GANs), the proposed GAN learns to generate image background and foregrounds separately and recursively, and stitch the foregrounds on the background in a contextually relevant manner to produce a complete natural image. For each foreground, the model learns to generate its appearance, shape and pose. The whole model is unsupervised, and is trained in an end-to-end manner with gradient descent methods. The experiments demonstrate that LR-GAN can generate more natural images with objects that are more human recognizable than DCGAN.