Geo-targeted image generation: Introducing location-specific datasets for fine-tuning stable diffusion models in India

Neeraj Kumar, Tanmay Singh, Amit Verma, Vivek Jain

In this paper, we introduce a novel approach for geo-targeted image generation by fine-tuning Stable Diffusion models using location-specific datasets in India. Addressing the challenge of producing culturally and geographically nuanced visual content, we focus on developing a specialized dataset for Maharashtra, enhancing the model's ability to generate images that authentically reflect the region's distinctive cultural and environmental landscapes. We detail the assembly and preparation of a dataset containing 282 high-quality images, each accompanied by rich contextual descriptions. This dataset forms the basis for adapting the “stabilityai/stable-diffusion-xl-base-1.0” model using a methodical fine-tuning process that integrates the Learning Rate Annealing (LoRA) technique. Our methodology ensures that the adapted model not only generates images with high fidelity but also captures the essence of local aesthetics and landmarks, thus significantly improving the relevance and personalization of generated images for applications in tourism, advertising, and cultural representation. Through our research, we aim to bridge the gap in performance of generative AI models across different geographies, providing a pathway towards more inclusive and representative AI applications that honor and reflect the diversity of human cultures.