Denoising diffusion probabilistic models in action: classifier-free vs classifier-guidance diffusion model approaches for Minecraft

While existing generative models have demonstrated success in producing high-quality images, modeling complex, multi-object Minecraft scenes remains challenging, particularly under varying lighting conditions. Traditional methods often decompose scenes into individual objects, but they struggle with accurately manipulating and arranging these objects in diverse lighting scenarios. In this paper, we explore how Classifier-Guided DDPM and Classifier-Free DDPM perform in control over specific object categories through conditional probabilities in minecraft scenes. We examine the impact of model architecture and parameter tuning, showing that while the Classifier-Guided DDPM achieves higher alignment with real data distributions, and faster training ,it more sensitive to parameter adjustments and relatively not focusing in structural coherence. In contrast, the Classifier-Free DDPM excels in structural coherence and generating more creative, finer local structure and details in scenes, with lower sensitive to parameter adjustments but higher training times.