Yu Zeng, Charles Ochoa, Mingyuan Zhou, Vishal M. Patel, Vitor Guizilini, Rowan McAllister
Phase-Preserving Diffusion (φ-PD) enables structure-aligned generation by preserving input phase in the diffusion process, improving spatial consistency in tasks like image-to-image translation.
The paper introduces a new method called Phase-Preserving Diffusion (φ-PD) that improves the generation of images and videos by maintaining their spatial structure. Traditional diffusion methods often disrupt the spatial arrangement of an image, which can be a problem for tasks that need geometric consistency. φ-PD addresses this by preserving the phase of the input data while allowing the magnitude to vary, leading to better alignment in the output. This method can enhance applications such as re-rendering and simulation enhancement, and it significantly improves performance in systems like the CARLA driving simulator.
