This work studies the problem of generating intermediate images from image pairs with large motion while maintaining semantic consistency. Existing methods either limit to small motion or focus on topologically similar objects, leading to artifacts and inconsistency in the