Diffusion models outperform GANs in image synthesis, achieving higher sample quality and better distribution coverage. The paper presents improvements in model architecture and introduces classifier guidance to enhance sample quality by trading off diversity for fidelity. The improved diffusion model achieves FID scores of 2.97 on ImageNet 128×128, 4.59 on 256×256, and 7.72 on 512×512. With classifier guidance, the FID scores improve to 3.94 on 256×256 and 3.85 on 512×512. The model also performs well on conditional image synthesis, achieving state-of-the-art results with as few as 25 forward passes per sample. The paper also compares the model with upsampling stacks, finding that combining both approaches yields the best results. The study shows that diffusion models can match or surpass GANs in sample quality while maintaining better distribution coverage. The results demonstrate that diffusion models are a promising direction for generative modeling, with potential for further improvements in sampling speed and performance on high-resolution image synthesis.Diffusion models outperform GANs in image synthesis, achieving higher sample quality and better distribution coverage. The paper presents improvements in model architecture and introduces classifier guidance to enhance sample quality by trading off diversity for fidelity. The improved diffusion model achieves FID scores of 2.97 on ImageNet 128×128, 4.59 on 256×256, and 7.72 on 512×512. With classifier guidance, the FID scores improve to 3.94 on 256×256 and 3.85 on 512×512. The model also performs well on conditional image synthesis, achieving state-of-the-art results with as few as 25 forward passes per sample. The paper also compares the model with upsampling stacks, finding that combining both approaches yields the best results. The study shows that diffusion models can match or surpass GANs in sample quality while maintaining better distribution coverage. The results demonstrate that diffusion models are a promising direction for generative modeling, with potential for further improvements in sampling speed and performance on high-resolution image synthesis.