Paper ID: 2412.17042

Adapting Image-to-Video Diffusion Models for Large-Motion Frame Interpolation

Luoxu Jin, Hiroshi Watanabe

The development of video generation models has advanced significantly in recent years. For video frame interpolation, we adopt a pre-trained large-scale image-to-video diffusion model. To enable this adaptation, we propose a conditional encoder, which serves as a simple yet effective trainable module. By leveraging the first and last frames, we extract spatial and temporal features and input them into the conditional encoder. The computed features of the conditional encoder guide the video diffusion model in generating keyframe-guided video sequences. Our method demonstrates superior performance on the Fr\'echet Video Distance (FVD) metric compared to previous deterministic approaches in handling large-motion cases, highlighting advancements in generative-based methodologies.

Submitted: Dec 22, 2024