Att-BiLSTM-GAN: A Temporal Coherence-Preserving GAN Framework for Dynamic Art Video Stylization

Abstract

Dynamic art video generation has emerged as a significant research area in computer vision and digital creativity, enabling the transformation of ordinary videos into visually compelling artistic content. However, existing methods often struggle to maintain temporal coherence, preserve motion integrity, and ensure faithful style transfer across consecutive frames, leading to artifacts such as flickering and content leakage. Specifically, an Attention-enriched Bidirectional Long Short-Term Memory integrated with Generative Adversarial Networks (Att-BiLSTM-GAN) method is introduced for dynamic art video generation. Dynamic action painting styles data are collected, featuring bold brushstrokes, splashes, and dynamic motion-like artistic patterns. Histogram Equalization (HE) is applied during preprocessing to normalize illumination and enhance contrast, while a Visual Geometry Group 16 (VGG16)-based encoder extracted spatial and textural features from content and style references. GAN-based style transfer techniques are applied to impose artistic attributes from reference artworks, preserving both global style patterns and local texture details. The incorporation of BiLSTM reduces temporal distortions by modeling frame-to-frame dependencies, while adaptive style loss functions balance stylistic richness. Experimental evaluations are conducted on the Dynamic Action Painting Styles dataset, with state-of-the-art baselines such as StyleMaster, SRCNN, VDSR, SRResNet, EDSR, TT-VSR, and baseline GAN models, the proposed Att-BiLSTM-GAN achieved superior results with PSNR (36.21), SSIM (0.96), CSD-Score (0.94), CLIP-Text (0.88), Motion Smooth (0.91), and FID (35.42), confirming significant performance gains and improved temporal coherence across all evaluation metrics. This research highlights the potential of combining sequential learning with style transfer for generating high-quality dynamic art videos.