Shulou(Shulou.com)11/24 Report--

CTOnews.com, November 14 (Xinhua)-- Google, in conjunction with the University of California, Berkeley (UC Berkeley), has developed a new generative AI method, idempotent generating network (IGN), which can replace the diffusion model (Diffusion Models).

Including generating antagonistic Network (GAN), Diffusion Model (Diffusion Models) and consistency Model (Consistency Models) released by OpenAI in March this year, the current mainstream generative AI models are random noises, sketches, or low-resolution or other corrupted images, which are mapped to the output (usually natural images) corresponding to a given target data distribution to generate images.

Taking the diffusion model as an example, we learn the target data distribution during training, and then perform "denoising" processing through multiple steps.

The Google research team has come up with a new generation model called idempotent generation network (IGN), which generates the right image from any form of input, ideally in one step.

This model can be thought of as a "global projector" that projects any input data onto the target data distribution. Unlike other existing model algorithms, it is not limited to specific inputs.

There are two main differences between IGN and GAN and diffusion model.

GAN requires a separate generator and discriminator model, while IGN is "self-antagonistic" and can play both roles.

The diffusion model requires incremental steps, while IGN can map input to data distribution in a single step.

The researchers demonstrated the potential of IGN using MNIST and CelebA datasets. The team demonstrated applications such as converting sketches into realistic images, generating images from noise, or repairing damaged images.

CTOnews.com encloses the detailed address of the paper here, which can be read in depth by interested users.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.