Although our method is designed to faithfully adhere to the user-defined animation, the user still has the option to use more diffusion motion prior by adjusting the hyperparameters in video interpolation.

Comparison with DiT-based methods, Diffusion-as-Shader and Go-with-the-Flow. Both methods struggle with fine-grained control such as object rotation.

Our framework is training-free and adaptable to the latest models. Results demonstrate improved details and more dynamic motion.

Experiments on error handling by reduced guidance. Our generated result corrects errors.