ObjBlur: A Curriculum Learning Approach With Progressive Object-Level Blurring for Improved Layout-to-Image Generation
Abstract: We present ObjBlur, a novel curriculum learning approach to improve layout-to-image generation models, where the task is to produce realistic images from layouts composed of boxes and labels. Our method is based on progressive object-level blurring, which effectively stabilizes training and enhances the quality of generated images. This curriculum learning strategy systematically applies varying degrees of blurring to individual objects or the background during training, starting from strong blurring to progressively cleaner images. Our findings reveal that this approach yields significant performance improvements, stabilized training, smoother convergence, and reduced variance between multiple runs. Moreover, our technique demonstrates its versatility by being compatible with generative adversarial networks and diffusion models, underlining its applicability across various generative modeling paradigms. With ObjBlur, we reach new state-of-the-art results on the complex COCO and Visual Genome datasets.
Primary Subject Area: [Generation] Generative Multimedia
Secondary Subject Area: [Content] Media Interpretation
Relevance To Conference: This paper introduces ObjBlur, a new approach to layout-to-image generation that utilizes curriculum learning by applying progressive object-level blurring. Blurring is a natural image degradation operation because low frequencies are retained over higher frequencies.
Strong blurring removes high-frequency details, resulting in a simpler signal without affecting the structural content of the image. Decreasing the blur strength produces a more complex signal with high-frequency details, thus exposing the model to a more difficult task. Therefore, blurring offers an intuitive and powerful approach to incrementally adjust task difficulty, ensuring a smooth training progression.
Supplementary Material: zip
Submission Number: 1720
Loading