emilwallner/Screenshot-to-code

Uses a CNN-RNN encoder-decoder architecture with GRU layers to convert UI screenshots into Bootstrap HTML/CSS, achieving 97% accuracy through a domain-specific token vocabulary of 16 elements rather than raw markup generation. Three model variants demonstrate the progression from proof-of-concept to production—the Bootstrap version generalizes effectively on new designs and trains on multiple GPUs, while the raw HTML variant remains experimental. Built on Keras/TensorFlow with pre-trained weights available, adapting the pix2code framework for practical UI-to-code synthesis.

16,515 stars. No commits in the last 6 months.