Stable Diffusion for Down Syndrome Face Generation

The growing need for diverse and inclusive datasets in computer vision has prompted research into synthetic image generation for underrepresented groups. This paper explores using Stable Diffusion XL (SD-XL 1.0-base) to generate highquality synthetic images of individuals with Down syndrome. By leveraging Low-Rank Adaptation (LoRA) and DreamBooth for concept-specific adjustments, our approach accurately reproduces key facial features. A curated dataset of 26 highresolution images was used to fine-tune the model. By training a CNN-based classification system using synthetic data, we demonstrate ResNet152V2 achieves 98.33% accuracy, underscoring the viability of synthetic images for data processing applications. The results enhance representation of individuals with Down syndrome in computer vision while addressing data scarcity and privacy concerns in biomedical imaging.