Cross-modal GNN–GRU framework for keyframe extraction and semantic-driven special-effects generation in animation videos

To address keyframe misidentification, semantic drift, and latency in animation processing, this study presents a cross-modal GNN–GRU framework for animation-video keyframe extraction and semantic-driven special-effects generation. The framework does not treat CNN, GNN, GRU, Wav2Vec, and attention fusion as isolated components, but couples them through frame-level graph construction, adaptive cross-modal weighting, confidence-gated modality compensation, and rhythm-aware effect timing control. Visual frames are encoded by a multi-scale CNN, speech semantics by a Wav2Vec Transformer, and rhythm signals by Fourier analysis. Experiments on an 18-hour multimodal anime dataset containing 13,600 frame-level samples demonstrate 91.7% keyframe accuracy, 88.9 semantic consistency, 1.4 s end-to-end latency, and 62.3 FPS throughput under GPU deployment. Additional comparisons with representative video summarization and multimodal baselines, together with ablation, complexity, perceptual-quality, and out-of-distribution tests, show that the proposed framework improves semantic alignment and real-time response while maintaining moderate computational cost. The derived feature tensors, frame-level labels, annotation guidelines, split files, and evaluation scripts are prepared for public release to improve reproducibility.