An Integrated Framework for Webpage Designing: A Comprehensive Study on Literature Review and Hybrid VAE-GAN

Abstract

The digital world is rapidly expanding with over 1.1 billion active websites. It creates a critical bottleneck for web development and remains intensive for non-technical users. The existing Artificial Intelligence (AI) and Machine Learning (ML) solutions offer automation but they frequently suffer from code hallucination and structural inconsistency. This research proposes an integrated hybrid framework design to bridge this gap by a visual perception with linguistic generation. The framework uses a multi-phase pipeline where a Variational Autoencoder (VAE) extracts UI features into a 512-dimensional probabilistic latent space (). This latent space is shared with a Generative Adversarial Network (GAN) Discriminator for visual realism. A transformer-based decoder performs the synthesis to fine-tune using GPT-2 tokenizer. This translates the latent design structure into sequential HTML code. The transformer can generate scattered tokens, causal making and adaptive repetition penalty will ensure the syntactic integrity for token generation for optimisation layer. The model is validated over a subset of 50,000 image-code pair entries sliced from the Web2Code dataset on the NVIDIA 4070 GPU over 8 GB VRAM. The 50-epoch model training experimental results demonstrate a significant and notable achievement of 100% syntax validity initialised on epoch 20 and remains consistent till the end with 0.2412 code generation loss and 0.2839 VAE structural loss. A comparative analysis shows that the framework achieves a superior design quality score of 9.2/10 and code quality of 9.4/10. The performance metrics further reveal a reduced load time of 1.4 seconds. Ultimately, the research provides a robust and scalable solution for automated webpage generation and moving toward an intelligent ecosystem that can combine human creativity with computational precision