Bridging the Gap: Consistent Image Outpainting via Training-Free Noise Optimization

Image outpainting has drawn increasing demands from many real-world applications. The core capacity called for this task is to generate image content beyond the boundaries that are semantically aligned with the source image. Compared to other image generation tasks, image outpainting remains very challenging since we need to identify the scene of the source image and generate new yet consistent boundaries with few local context. However, one common propensity for the outpainting techniques is to generate irregular high-frequency patterns. Furthermore, the dominating data-driven learning paradigm utilized by the existing state-of-the-art methods would require sophisticated model design, significant computation cost and introduce potential bias as well.