A Framework with Multi-Scale Hybrid Mamba Voxel Flow for Video Prediction

Video prediction is a critical task in video processing and generation, with far-reaching implications for various downstream applications. However, existing methods often produce blurred predicted frames and fail to maintain structural continuity in objects. To address these challenges, we propose a Multi-Scale Hybrid Mamba Voxel Flow framework that employs a progressive refinement strategy in combination with adaptive feature extraction modules. The framework begins by generating coarse optical flow estimates and predicted frames, which are progressively refined at lower resolutions to enhance detail and ensure temporal coherence. Specifically, Mamba Blocks are designed to capture complex global motion patterns, while Spatial Aggregation Blocks aggregate spatial context across different scales. Simam Modules further enhance feature representation by selectively focusing on significant spatial regions. Additionally, multi-level residual connections and depthwise channel separations are incorporated to reduce computational complexity. Experimental results show that the proposed method significantly improves the clarity and spatial consistency of predicted frames, outperforming state-of-the-art techniques.