Optimization of automatic generation and light rendering of film and television scenes based on generative adversarial network and multimodal feature fusion algorithm

An optimization model for automatic generation and light and shadow rendering of film and television scenes based on generative adversarial network and multimodal feature fusion is proposed in this paper. Based on the traditional GAN architecture, the model introduces semantic feature extraction and cross-modal mapping mechanism to unify text, image and audio features into a high-dimensional semantic space, and realizes the scene construction through the double constraints of generator and discriminator. In order to enhance the realism, the light and shadow effects are optimized by combining the physical lighting model and the neural rendering method. The experimental results show that the FID value of this method on MovieNet and COCO-Scene datasets is reduced to 28.3, which is significantly better than 65.4 of GAN and 72.1 of VAE; the IS reaches 4.72, which is significantly improved compared with the 3.95 of the multimodal GAN, and the PSNR and SSIM reach 28.1 dB and 0.89, which are comprehensively better than the comparison methods. The results validate the comprehensive advantages of the model in terms of image fidelity, semantic consistency and light and shadow performance.