In the framework of the recently held International Machine Learning Conference (International Conference on Machine Learning, icml), a new revolutionary artificial intelligence system called Blackout Diffusion was presented. This technology allows for the generation of images from an absolutely empty image, setting it apart from other existing generative models like Dall-E or Midjourney. Notably, Blackout Diffusion does not require any initial data to initiate the generation process.
Javier Santos, an AI researcher from the National Laboratory of Los Alamos and co-author of the work, stated that generative models are igniting a new industrial revolution by automating tasks such as code generation, legal document creation, and even artwork.
One of Blackout Diffusion’s significant achievements is its ability to operate in discrete spaces, unlike existing models that function in continuous spaces. This expands the potential for the technology’s applicability in scientific research and other fields.
The process of generating images through Blackout Diffusion involves diffusion models that create samples similar to the data they study. These models work by taking an image and repeatedly adding noise until it becomes unrecognizable. Throughout this process, the model learns to return the image to its original state. Unlike current models, Blackout Diffusion does not require input data to initiate image creation.
The project leader of Blackout Diffusion, physicist Ian-Ting Lin from Los Alamos, claims that the quality of images created using the system is comparable to the results of current models while requiring fewer computing resources. The team tested the technology on several standard data sets that include the Database of the National Institute of Standards and Technologies (National Institute of Standards and Technology, NIST), CIFAR-10 data containing images of objects from 10 different classes, and a set of attributes of Celebface, which consists of over 200,000 images of human faces.
The key difference between discrete and continuous spaces lies in the limited nature of values in the former, whereas the latter allows values to exist within a certain spectrum without limitation. Discrete spaces contain separated values, such as the number of people in a room or categories of colors, while continuous spaces enable values to change continuously and occupy any point within a range, such as temperature measurements.