以 Mamba 為代表的狀態空間模型(SSM)已成為 Transformer 在長程序列建模上的重要替代方案，然而將特徵圖序列化為一維序列會削弱二維空間的局部性，而這正是高解析度影片插幀(VFI)中描繪清晰運動邊界所不可或缺的。本論文提出HMA-VFI一個用於 VFI 的混合式 Mamba–Attention 框架，在共享的區域-全域協同主幹(Local-Global Synergistic backbone)中，將線性複雜度的 Mamba2D 掃描與局部視窗式跨幀注意力(IFA)相結合，以同時建模全幀全域脈絡與局部跨幀對應關係。該框架進一步整合 Flow-Aligned Cross Frame Fusion(FA-CFF)以進行預對齊的時序特徵聚合、一個由粗到細的光流估計器，以及配備信心門控殘差線索(confidence-gated ErrCue)的 U-Net 細化解碼器。模型僅在 Vimeo90K 上訓練，並在 Vimeo90K、SNU-FILM 與 Xiph 三個基準上不使用測試時增強進行評估;在 Vimeo90K 與 SNU-FILM 上達到具競爭力的效果，並在本文比較中取得最佳的 Xiph 4K cropped 成績，包含對使用額外訓練資料之方法的比較。上述結果驗證了所提框架對大運動量、高解析度影片插幀的有效性。;State-space models (SSMs) such as Mamba have emerged as important alternatives to Transformers for long-range sequence modeling， but serializing feature maps into one-dimensional sequences weakens the two-dimensional spatial locality essential for preserving motion boundaries in high-resolution video frame interpolation (VFI). In this thesis， we propose HMA-VFI， a hybrid Mamba–Attention framework for VFI that combines linear-complexity Mamba2D scanning with window-based Inter-Frame Attention (IFA) in a shared Local-Global Synergistic backbone， enabling joint modeling of full-frame global context and local cross-frame correspondences. The framework further incorporates Flow-Aligned Cross-Frame Fusion (FA-CFF) for pre-aligned temporal feature aggregation， a coarse-to-fine flow estimator， and a U-Net RefineNet guided by a confidence-gated ErrCue. Trained exclusively on Vimeo90K and evaluated on Vimeo90K， SNU-FILM， and Xiph without test-time augmentation， the model achieves competitive performance on Vimeo90K and SNU-FILM， and records the best Xiph 4K cropped results in our comparison， even against methods trained on additional data. These results validate the effectiveness of the proposed framework for large-motion， high-resolution video frame interpolation.
