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

5 Mar 2017  ·  Jianwei Yang, Anitha Kannan, Dhruv Batra, Devi Parikh ·

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.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Conditional Image Generation CIFAR-10 LR-GAN Inception score 7.17 # 16
Image Generation CIFAR-10 LR-GAN Inception score 7.17 # 63
Image Generation CUB 128 x 128 LR-GAN FID 34.91 # 4
Inception score 13.50 # 3
Image Generation Stanford Cars LR-GAN FID 88.80 # 4
Inception score 5.25 # 3

Results from Other Papers


Task Dataset Model Metric Name Metric Value Rank Source Paper Compare
Image Generation Stanford Dogs LR-GAN FID 54.91 # 4
Inception score 10.22 # 3

Methods