Comparative Analysis of DDPM and DDIM Sampling Methods in Conditional Generation Models Using Minecraft Dataset

This study systematically compares the performance of the two sampling methods of DDPM and DDIM and the combination of unet, cspdarkunst and unetv2 networks in the conditional generation model by using the "Minecraft" dataset containing forest, sea and desert scenes. Experimental results show that the image quality differences between DDPM and DDIM are minimal, the Unetv2 network outperforms in terms of color accuracy, clarity, and scene complexity. Further experiments explore the effects of learning rate, activation functions, and attention mechanisms on image generation. The experimental results show that setting the learning rate to 3e-5 achieves the best image generation quality, and adding attention mechanisms enhances image realism, particularly in desert scenes.