New Diffusion Models Surpass GANs in High-Resolution Image Synthesis
Researchers demonstrate that diffusion models achieve state-of-the-art image sample quality, outperforming GANs by using improved architectures and classifier guidance. The models achieve record-breaking scores across various ImageNet resolutions while maintaining excellent distribution coverage.
Diffusion models officially surpass GANs in generating high-quality images, according to a groundbreaking new paper. Researchers achieve this state-of-the-art performance in unconditional image synthesis by designing a better neural network architecture through careful testing and elimination of unnecessary components.
For conditional image generation, the team introduces classifier guidance, a simple and compute-efficient technique that uses gradients from a separate classifier to trade diversity for image fidelity. This approach yields impressive FID scores of 2.97 on ImageNet 128x128, 4.59 on ImageNet 256x256, and 7.72 on ImageNet 512x512, matching the performance of BigGAN-deep with significantly fewer computational steps.
The researchers also find that combining this classifier guidance with upsampling diffusion models pushes the results even further. This combination improves the FID scores to 3.94 on ImageNet 256x256 and 3.85 on ImageNet 512x512, all while ensuring better coverage of the overall data distribution than previous generative models.