FUN-VFI: Flow-Upscaling Network for 4K Video Frame Interpolation

4K video frame interpolation is challenging due to large motions and unexpected occlusion between two consecutive frames. To address this challenge, we propose the Flow-Upscaling Network for 4K Video Frame Interpolation, denoted as FUN-VFI. Specifically, it consists of two main modules: global flow field estimation and adjustable upscaler. In global flow field estimation, the flow estimator predicts bilateral flow fields between two downscaled consecutive frames. Subsequently, we iteratively enhance these bilateral flow fields using the adjustable upscaler. The core of the upscaler is the ConvNext block, a pure CNN model that can rival vision transformers while preserving the simple structure and efficient runtime of convolutional neural networks. Last, we warp two consecutive input frames using refined flow fields and blend them to generate the reconstructed intermediate frame. Experimental results demonstrate that the proposed method achieves outstanding performance across diverse benchmark datasets. Furthermore, compared with existing methods like M2M-PWC and BI-Former, FUN-VFI accelerates inference times by approximately 2-6 times at 4K resolution, while consistently delivering superior outcomes.