[Paper Review] Fast-converging Conditional Generative Adversarial Networks for Image Synthesis

Chengcheng Li, Zi Wang+1 more|arXiv (Cornell University)|May 5, 2018

Generative Adversarial Networks and Image Synthesis6 citations

TL;DR

This paper proposes Fast-converging Conditional GAN (FC-GAN), which accelerates training by introducing an advanced auxiliary classifier that distinguishes real classes from a dedicated 'fake' class, avoiding confusion between real and generated samples. The method achieves faster convergence and competitive image quality compared to AC-GAN, with training reaching equilibrium up to 30% faster on MNIST and CIFAR-10 datasets.

ABSTRACT

Building on top of the success of generative adversarial networks (GANs), conditional GANs attempt to better direct the data generation process by conditioning with certain additional information. Inspired by the most recent AC-GAN, in this paper we propose a fast-converging conditional GAN (FC-GAN). In addition to the real/fake classifier used in vanilla GANs, our discriminator has an advanced auxiliary classifier which distinguishes each real class from an extra `fake' class. The `fake' class avoids mixing generated data with real data, which can potentially confuse the classification of real data as AC-GAN does, and makes the advanced auxiliary classifier behave as another real/fake classifier. As a result, FC-GAN can accelerate the process of differentiation of all classes, thus boost the convergence speed. Experimental results on image synthesis demonstrate our model is competitive in the quality of images generated while achieving a faster convergence rate.

Motivation & Objective

• To address the slow convergence of conditional GANs, particularly in distinguishing real and generated samples across multiple classes.
• To improve the stability and speed of training by redesigning the auxiliary classifier in the discriminator to avoid mixing real and generated data in the same class.
• To enhance the discriminator's ability to learn class-specific features and real/fake distinctions simultaneously.
• To achieve faster convergence without sacrificing image quality in conditional image synthesis tasks.
• To validate the method on standard benchmarks such as MNIST and CIFAR-10 with quantitative and qualitative evaluation.

Proposed method

• The discriminator in FC-GAN includes an advanced auxiliary classifier that assigns real samples to their true classes and generated samples to a dedicated 'fake' class, distinct from real classes.
• This design prevents the auxiliary classifier from being confused by mixing real and generated data in the same class, as in AC-GAN, thereby improving class discrimination.
• The 'fake' class also enables the auxiliary classifier to function as a secondary real/fake discriminator, reinforcing the main discriminator's objective.
• The overall loss combines the standard GAN loss (source loss) and the classification loss over N+1 classes (N real classes + 1 fake class).
• The generator is conditioned on both the latent noise and the class label, ensuring class-controlled image generation.
• The architecture uses convolutional and deconvolutional layers with batch normalization and ReLU activations, as detailed in Table 1 for MNIST.

Experimental results

Research questions

• RQ1 Can introducing a dedicated 'fake' class in the auxiliary classifier improve convergence speed in conditional GANs?
• RQ2 Does separating real and generated data in the classification head reduce confusion and improve training stability?
• RQ3 Can the auxiliary classifier simultaneously serve as a real/fake discriminator, enhancing overall GAN training dynamics?
• RQ4 How does FC-GAN compare to AC-GAN in terms of convergence speed and image quality on standard benchmarks?
• RQ5 Does the proposed method maintain or improve image fidelity and diversity while accelerating training?

Key findings

• FC-GAN achieves convergence in approximately 13 epochs on MNIST, compared to 20 epochs for AC-GAN, indicating a significant speedup in training.
• On CIFAR-10, FC-GAN reaches comparable performance to AC-GAN in 20 epochs, while AC-GAN requires 50 or more epochs to match the results.
• The Parzen window estimate improved from 581.8 ± 5.4 (AC-GAN) to 646.9 ± 5.3 (FC-GAN) on CIFAR-10, indicating better density estimation.
• The inception score on CIFAR-10 increased from 4.190 ± 0.08 (AC-GAN) to 6.360 ± 0.14 (FC-GAN), showing improved image quality and diversity.
• On MNIST, FC-GAN achieved a slightly higher inception score (2.238 ± 0.03) than AC-GAN (2.216 ± 0.04), with minimal improvement in Parzen window estimate.
• Visual samples show that FC-GAN generates sharper and more diverse digits after only 10 epochs, while AC-GAN requires 20+ epochs to reach similar quality.