LLM-DiffAug: Enhancing few-shot object detection via LLM-Guided diffusion augmentation

Yunqing Jiang, Sunyuan Qiang, Wuchen Li, Yanyan Liang

Abstract: Highlights•LLM-guided Prompt Generation: We propose LLM-guided Diffusion models for Data Augmentation (LLM-DiffAug), which leverages LLM to generate diverse prompts for diffusion model inpainting.•Inpainting Alignment: We adopt inpaint alignment to reduce background influence and provide more precise control over inpainted areas.•Bounding Box Constraint: We implement box constraints on inpainted objects to better align with given bounding boxes.•Performance Achievement: Systematic evaluation shows significant improvements over baselines and state-of-the-art methods.

External IDs:dblp:journals/kbs/JiangQLL25