When synthetic plants get sick: Disease graded image datasets by novel regression-conditional diffusion models

This paper introduces DiffusionPix2Pix, an innovative extension of diffusion models (DMs) that revolutionizes synthetic image generation by seamlessly integrating image priors, surpassing existing state-of-the-art models. Key contributions include regression (graded) conditioning and an arbitrary binary mask, enabling regression-conditional image-to-image translation. DiffusionPix2Pix is compared with Pix2Pix-G and Pix2Pix-GD, two alternative models that rely on image-conditioned GANs adapted for an additional regression conditional task. The model is applied to generate a graded plant disease dataset focusing on Puccinia striiformis symptoms, using disease degree as an additional conditioning input to control the level of disease in generated images. Experiments demonstrate that DiffusionPix2Pix outperforms GAN-based approaches in both sample fidelity and diversity, achieving an Improved Precision (fidelity) of 0.81 (versus 0.45 and 0.47) and an Improved Recall (diversity) of 0.58 (versus 0.31 and 0.31). Furthermore, DiffusionPix2Pix obtained the best Fréchet Inception Distance (FID), with a score of 31.61 compared to 57.38 and 54.34 for GAN-based models. Additionally, perception-based tests with field technicians showed 71.3% of images generated by DiffusionPix2Pix were classified as authentic, significantly outperforming the 20.19% and 22.22% rates for GAN-based models. These findings substantiate the performance of the proposed DiffusionPix2Pix model, both quantitatively and through subjective assessments by domain experts, highlighting its potential in applications requiring precise regression conditioning. • A novel diffusion-based method for synthesizing diseased plant leaf images. • The method provides precise control over leaf shape and disease level. • DiffusionPix2Pix outperforms GAN-based approaches with higher quality images. • Experts note more realistic images with DiffusionPix2Pix vs. GAN-based methods.