Synthetic Data for PET and HDPE Classification: A Comparison of Diffusion and Autoregressive Models

Marine plastic pollution poses a significant environmental threat, with polyethylene terephthalate (PET) and high-density polyethylene (HDPE) debris among the most prevalent yet difficult to detect in complex underwater environments. Deep learning models for automated detection are often limited by scarce, imbalanced, or hard-to-collect underwater datasets. This study investigates the use of state-of-the-art generative artificial intelligence models to create synthetic underwater images of plastic debris. Two paradigms were compared: a transformer-based autoregressive model (DALLE-3) and a denoising diffusion model (Stable Diffusion XL, SDXL). A structured prompt engineering workflow was developed to encode object features, environmental context, spectral cues, and biotic interactions.