Progressive Outpainting Task
This project was inspired by stablediffusion-infinity. When performing outpainting based on the progressive approach with Latent Diffusion Models, it is difficult to preserve the overall context and style of the original input as the number of generations increases. To avoid this, if outpainting is performed within a single window, it becomes challenging to achieve high-resolution outpainting and blurring may occur. To overcome these limitations, we propose a novel progressive outpainting method that also considers a global window. The module that considers the global window is connected to the frozen pre-trained denoiser and trained, which reduces training time while preserving the performance of the pre-trained denoiser. Our perspective is fundamentally different from conventional reference-based painting approaches that explicitly inject objects or styles from reference images. Even if there is no corresponding object to be generated, our method infers plausible composition automatically without explicit information. For more details, please refer to the paper and supplementary materials.
