A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a significant advancement in AI-powered image generation. The research focuses on improving the spatial understanding of diffusion models, which are commonly used in tools like DALL-E and Stable Diffusion. These models often struggle with accurately representing the spatial relationships between objects in a scene, such as placing a cat correctly to the left of a dog. The MIT team developed a training technique that feeds the model additional data about the bounding boxes and spatial locations of objects within an image. This method, tested on models trained from scratch and fine-tuned on existing ones, resulted in generated images that more faithfully followed spatial guidance. The improvement was notable, with the fine-tuned model showing a 30% increase in spatial accuracy compared to its base version. This research addresses a core limitation in current text-to-image AI, paving the way for more reliable and controllable generation for applications in design, gaming, and robotics. Read the full article at: https://technologyreview.com/2024/07/10/1094825/mit-ai-image-generation-spatial-reasoning/