IG03 - Research Group for Computer Science and Human-Computer Interaction

Research on generative models for automatic synthesis of images with applications in data augmentation and workflow automation. We develop methodologies for fine-tuning diffusion models (Stable Diffusion) on small domain-specific datasets and their application for generating diverse scenarios (weather conditions, lighting) not represented in original data. We investigate automated pipelines combining large language models for prompt generation with image synthesis models (SDXL, Flux, Midjourney), and conduct systematic evaluations of generated content quality. Applications include data augmentation for crosswalk segmentation in assistive systems for visually impaired individuals and automatic image generation for agricultural news articles.

Technologies Covered

• Diffusion models (Stable Diffusion, SDXL, SD 3.5, Flux, Midjourney)
• Fine-tuning diffusion models on small datasets
• LLM-based prompt generation (GPT, Gemini, Claude)
• Synthetic data augmentation for semantic segmentation
• Conditional image generation (weather, lighting variations)
• Human evaluation methodologies for generative AI
• Synthetic-real dataset creation and benchmarking