Enhancing weed detection in strawberry plasticulture: Evaluating stable diffusion models for synthetic image generation of Geranium carolinianum

The development of vision-guided smart sprayers depends heavily on diverse and high-quality datasets, either collected in the field or sourced from public repositories. However, for certain crops, acquiring such datasets is both labor-intensive and challenging. Text-to-image generative models, such as those based on Stable Diffusion (SD), offer a promising alternative by generating synthetic images from a limited number of real samples. This study explores the potential of two SD models (SDXL and SD3.5L) to generate realistic images of Geranium carolinianum growing alongside strawberry plants. The goal was to assess the YOLOv11 model performance by evaluating training and validation strategies involving: 1) Real images only, 2) 50% Real + 50% SD3.5L, 3) 50% Real + 50% SDXL, 4) SD3.5L synthetic only, and 5) SDXL synthetic only. A total of 400 high-resolution field images were collected during the 2024-2025 strawberry season in Florida, with 50 representative images of G. carolinianum at various growth stages selected to fine-tune the SD models. The quality of generated images was evaluated using Fréchet Inception Distance (FID), Inception Score (IS), and Structural Similarity Index Measure (SSIM). YOLOv11 model performance was assessed using F1-scores and mean average precision at IoU thresholds of 50% to 95% (mAP50, mAP50-95%). SDXL achieved superior FID and IS scores and produced more photorealistic images, whereas SD3.5L generated images with greater morphological accuracy, variability, and higher SSIM values. YOLOv11 models trained on a mix of real and synthetic images performed comparably to models trained only on real data, with no statistically significant differences in detection accuracy. Models trained exclusively on synthetic data yielded lower performance. These results highlight the potential of generative models to supplement dataset creation, significantly reducing field collection efforts, particularly valuable during off-season periods, for species with limited geographic distribution, under environmental constraints, or for rare or newly discovered species.