Apple has recently open-sourced a brand-new image generation model along with its accompanying training method. This new achievement has garnered widespread attention in the field of image generation.
The model has shown remarkable performance, particularly in its powerful zero-shot generalization capabilities. It is noteworthy that it achieved such impressive results using only the CC12M dataset, which contains 12 million images. The ability to achieve such outstanding performance with a relatively small data scale undoubtedly proves the innovation and efficiency of this model and training method.
On the technical side, Apple has proposed a new diffusion process with unique advantages. This process can simultaneously denoise inputs at multiple resolutions. This feature allows the model to maintain good performance when processing images of different resolutions, enhancing the model's applicability and versatility.
Additionally, the model employs a novel Nested UNet architecture. In this architecture, there is a clever nesting relationship. Specifically, the features and parameters of smaller-scale inputs are carefully nested within those of larger-scale features and parameters. This nesting structure enables the model to better capture image features at different scales, thereby further enhancing the model's performance and providing strong assurance for the quality and efficiency of image generation.