PhysicEdit: Teaching Image Editing Models to Respect Physics
AI Summary: PhysicEdit is a novel framework for image editing that addresses the limitations of traditional instruction-based models by treating image modifications as physical state transitions rather than static transformations. The authors propose a new dataset, PhysicTran38K, which consists of approximately 38,000 video-instruction pairs that illustrate physical transitions across various domains, including mechanical and optical phenomena. This dataset enables the model to learn the intermediate steps of physical evolution, thereby improving the realism of edits in scenarios that require adherence to physical laws. By modeling edits as state evolution problems, PhysicEdit aims to enhance the accuracy of image modifications in physics-heavy contexts.