Training a 1.57B-Parameter Dreamer 4 World Model for Under $150

An independent developer successfully trained a 1.57B-parameter Dreamer 4 world model from scratch for under $150 using procedurally generated data. The project achieves a tokenizer PSNR of 40.41 and a Fréchet Video Distance (FVD) of 32.19, demonstrating that high-quality world models can be trained affordably. This project shows that training large-scale world models no longer requires the massive budget of frontier AI labs, democratizing access to advanced reinforcement learning research. By using procedurally generated data instead of scraped videos, developers can maintain precise control over action-response mapping in simulated environments. The model was trained on 9.6 million frames and can simulate up to 144 frames end-to-end before the generation degrades. The developer switched from Google's Genie architecture to Dreamer 4 because Genie's unsupervised action learning made the controls too loose for effective interaction.