"ControlThink-50K"

"ControlThink-50K is a high-quality visual reasoning dataset designed for controllable image generation. It focuses on helping multimodal large language models better understand control images, such as edge maps, depth maps, and segmentation maps. Each sample includes a control image, an original text prompt, a chain-of-thought reasoning process, and an enhanced prompt with richer semantic details. The dataset is mainly constructed with GPT-4o and further refined through rejection sampling and quality filtering to ensure reliable reasoning. It supports both supervised fine-tuning and reinforcement fine-tuning, enabling models to bridge the semantic gap between sparse prompts and target images, thus improving generation quality and semantic consistency. "