DDIM stands for Denoising Diffusion Implicit Models. It is a type of generative model used in image generation tasks. DDIMs are an extension of Denoising Diffusion Probabilistic Models (DDPMs), which are based on the idea of gradually adding noise to data and then learning to reverse this process to generate new data samples.
The key difference with DDIMs is that they introduce a non-Markovian diffusion process, which allows for a more efficient sampling process. This means that DDIMs can generate high-quality images with fewer steps compared to traditional DDPMs, making them faster and more computationally efficient while maintaining or even improving the quality of the generated images.
Referenced in:
All notes