DiT-DDIM: Invisible Diffusion Image Generation Model Based on Transformer Framework

Diffusion modeling is a powerful generative method, has been widely used in computer vision fields such as image synthesis, image restoration and many other tasks. The original diffusion model uses U-net network and combines with Markov chain to generate data, but with the development of the times, for the reduction of the image generation time and the quality of the image generation is more and more demanding, the traditional diffusion model needs to be improved in terms of the generation speed and the generation quality, so for this reason, we have proposed a new framework. By using the Transform framework model to learn the denoising process in a large amount of image data, and using the deterministic sampling mechanism of nonMarkov chain in the sampling process, so as to ensure the quality of image generation while improving the generation efficiency. In DiT-DDIM model with 20 step interval as the standard case single picture generation time is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 1}$</tex> seconds, and the best FID is 4.7. The research in this paper has certain advantages in terms of applicability and practicability, and provides new research ideas and methods for the direction of picture.