Research on Minecraft Image Generation Method Based on Denoising Diffusion Probabilistic Models

Generative models have shown great potential in game image generation, especially for pixel-style games like Minecraft, where automatic content generation can significantly enhance the player experience. This paper investigates the application effects of unconditional DDPM, conditional DDPM, classifier-free DDPM, and guided DDPM in Minecraft image generation based on the Denoising Diffusion Probabilistic Model (DDPM). First, we trained an unconditional DDPM using over 3,000 images extracted from Minecraft videos to demonstrate its basic performance in terms of generation quality and detail retention. Subsequently, through comparative analysis, we found that classifier-free DDPM performed best in terms of detail richness and diversity, while guided DDPM was more suitable for generating consistent style content. Further evaluation of image generation effectiveness using the Fréchet Inception Distance (FID) indicates that classifier-free DDPM has a significant advantage in generation quality.