UPHSID: Underwater PET–HDPE Synthetic Image Dataset

Dataset Description

This dataset contains a collection of synthetic underwater images of PET and HDPE plastic debris, generated using two state-of-the-art generative AI models: Stable Diffusion XL (SDXL) and DALL·E-3. The dataset was developed to address the scarcity of balanced, high-quality underwater imagery for deep learning models used in marine plastic detection.

All images were generated using a structured, multi-stage prompt engineering workflow described in the accompanying publication titled "Synthetic Data for PET and HDPE Classification: A Comparison of Diffusion and Autoregressive Models" at 33rd International Conference on Artificial Intelligence and Cognitive Science (AICS 2025). The image generation design was driven by prompt engineering, which followed a structured, iterative process (prompts P0-P6). The steps builds from basic objectdefinitions (P0) with progressivley scaling to include environmental and object specific features, such as environmental context (P1), optical/physical parameters (P2), material-specific tokens (P3), biotic interactions (P4), and spectral cues (P5). An optional final composite stage (P6) combined these elements into physics-compliant prompts that integrated materials, flora/fauna, and spectral descriptors. Negative prompts were systematically applied after P3–P5 to discard artefacts (e.g., logos, watermarks, non-plastic materials, and unrealistic lighting). 

Contents

The dataset includes 200 synthetic underwater images:

100 PET images

50 generated using SDXL

50 generated using DALL·E-3

100 HDPE images

50 generated using SDXL

50 generated using DALL·E-3

All images are provided in RGB format (.png/.jpg depending on model output).