A Survey of Training-free Diffusion-based Image Generation with Free-form Mask

Layout-to-image generation is a task that generates realistic images based on given layouts and corresponding textual descriptions. The layout provides structural information about the image, such as descriptions, positions, and sizes of objects. Traditional methods for layout-to-image generation relied on bounding boxes, which represent only fixed-form layouts. Recently, approaches using free-form masks have gained attention, as they enable more flexible control over the shapes and positions of objects. Among these, training-free methods have been proposed that leverage pre-trained diffusion models without additional training. These methods adjust modified attention and guidance mechanisms to steer the image generation process during the inference phase of the diffusion model. In this paper, we review training-free diffusion-based image generation methods that utilize free-form masks. We focus on three representative methods: Paint-with-Words, MultiDiffusion, and Zero-Painter. We analyze their generation strategies and key mechanisms, as well as their limitations regarding spatial accuracy and consistency in object placement.