PROCEEDINGS OF THE 2026 GENETIC AND EVOLUTIONARY COMPUTATION CONFERENCE, GECCO 2026 COMPANION
Abstract
While Latent Diffusion Models (LDMs) excel at synthetic data generation, they suffer from severe training instability on the small, homogeneous UAV datasets typical of precision agriculture. In this work, we investigate these architectural bottlenecks by training a Denoising Diffusion Probabilistic Model (DDPM) from scratch on constrained pineapple plantation imagery. Our systematic comparison reveals that continuous latent spaces (β-VAE) preserve complex crop structures, whereas discrete architectures (VQ-VAE) fail under scarce data conditions. Additionally, we demonstrate that replacing destabilizing global attention mechanisms with a purely convolutional U-Net backbone ensures smooth convergence in repetitive visual domains. Ultimately, combining this robust architecture with DDIM sampling yields a stable, highly efficient framework for low-data agricultural image synthesis. Code and models will be available at https://github.com/imagine-laboratory/Latent-Diffusion.