arxiv_ai 95% Match Research Paper AI Researchers,ML Engineers,Computer Vision Practitioners,GAN Developers 1 week ago

BOLT-GAN: Bayes-Optimal Loss for Stable GAN Training

generative-ai › gans

📄 Abstract

Abstract: We introduce BOLT-GAN, a simple yet effective modification of the WGAN framework inspired by the Bayes Optimal Learning Threshold (BOLT). We show that with a Lipschitz continuous discriminator, BOLT-GAN implicitly minimizes a different metric distance than the Earth Mover (Wasserstein) distance and achieves better training stability. Empirical evaluations on four standard image generation benchmarks (CIFAR-10, CelebA-64, LSUN Bedroom-64, and LSUN Church-64) show that BOLT-GAN consistently outperforms WGAN, achieving 10-60% lower Frechet Inception Distance (FID). Our results suggest that BOLT is a broadly applicable principle for enhancing GAN training.

Authors (5)

Mohammadreza Tavasoli Naeini

Ali Bereyhi

Morteza Noshad

Ben Liang

Alfred O. Hero III

Submitted

October 29, 2025

arXiv Category

cs.LG

arXiv PDF

Key Contributions

BOLT-GAN introduces a simple yet effective modification to the WGAN framework, inspired by the Bayes Optimal Learning Threshold (BOLT). This modification implicitly minimizes a different metric distance than the Wasserstein distance, leading to significantly improved training stability and better image generation quality, as evidenced by lower FID scores.

Business Value

Enables more reliable and efficient training of GANs for generating high-fidelity synthetic data, which can be used in various applications like data augmentation, content creation, and simulation.

Paper Metadata

Innovation Type

Algorithmic Modification

Deployment Feasibility

High, as it's a modification to an existing framework (WGAN) and aims for simplicity.

Limitations Addressed

Addresses the common issue of training instability in GANs, particularly WGANs, which can hinder the generation of high-quality images and reliable convergence.

Performance Gains

Achieves 10-60% lower Frechet Inception Distance (FID) compared to WGAN.

Technical Tags

GANsWGANBayes Optimal Learning Threshold (BOLT)Lipschitz continuitydiscriminatortraining stabilityimage generationFrechet Inception Distance (FID)

Research Topics

Generative Adversarial NetworksGAN Training StabilityImage GenerationDeep Learning Theory

Methods & Architectures

Modification of WGAN frameworkBayes Optimal Learning Threshold (BOLT) inspirationEmpirical EvaluationFID calculation Generative Adversarial Networks (GANs)Wasserstein GAN (WGAN)

Applications & Tasks

Image Generation Synthetic Data Generation GAN training instabilityMode collapseDifficulty in achieving stable convergence Stable GAN TrainingHigh-quality Image Generation

Datasets & Benchmarks

Datasets

CIFAR-10, CelebA-64, LSUN Bedroom-64, LSUN Church-64

Benchmarks

10-60% lower Frechet Inception Distance (FID)

Frechet Inception Distance (FID)

Related Fields

Generative AIDeep LearningComputer VisionMachine Learning Theory

Keywords

GANsWGANBOLT-GANtraining stabilityimage generationdiscriminatorLipschitz continuityFIDgenerative modelsdeep learning

Academic Context

#Generative Adversarial Networks#GAN Training Stability#Image Generation#Deep Learning Theory

Technology Stack

Frameworks & Libraries

WGAN

Commercial Potential

Potential Products

Improved GAN training librariesTools for generating high-quality synthetic imagesContent creation platforms

Target Industries

Media and EntertainmentGamingAdvertisingE-commerceHealthcare (synthetic data)

Use Case Examples

Generating realistic faces for virtual avatarsCreating synthetic datasets for training other AI modelsArtistic image generation

Competitive Edge

Offers a more stable and effective training method for GANs compared to standard WGAN, leading to better generation quality.

Market Opportunity

Large and growing market for generative AI and synthetic data.

Revenue Models

Licensing the technologyoffering services for synthetic data generation.

Resource Requirements

Compute Needs

Requires significant GPU resources for GAN training.

Data Requirements

Standard image datasets for training GANs.

Deployment Constraints

Computational cost of training GANs.

Scalability

Aims to improve the reliability and thus practical scalability of GAN training.

Production Readiness

Maturity Level

Research

Time to Market

1-2 years for integration into existing GAN frameworks.

Patent Potential

Moderate, for the BOLT-GAN modification and its application.

View Full Paper Back to Papers