One Pass Is Not Enough: Recursive Latent Refinement for Generative Models (arxiv.org)

A new paper presents Recursive Latent Refinement (RTM), a significant advancement in generative models that addresses persistent challenges in image generation, such as mode collapse and insufficient diversity. Traditional evaluation metrics like FID (Fréchet Inception Distance) often conflate sample quality with coverage, leading models to score well despite failing to truly represent the data distribution. RTM proposes an iterative latent mapping process instead of the single-pass approach used in many style-based generators, emphasizing the importance of maintaining high precision and recall alongside FID scores. This innovation is crucial for the AI/ML community as it not only enhances the quality of generated images but also promotes a more exhaustive coverage of data modes. By integrating RTM with Implicit Maximum Likelihood Estimation (IMLE), the approach achieves superior metrics across established datasets like CIFAR-10 and CelebA-HQ, showcasing improvements in both precision and coverage. The ability of RTM to boost performance even for widely adopted models such as StyleGAN2 highlights its potential for broader applications, making it a promising approach for tackling the ongoing limitations in generative model performance.