AI generates images 30 times faster with this MIT technique

AI generates images 30 times faster with this MIT technique

Midjourney or Stable Diffusion allows you to generate a large number of images very quickly from prompts. If the result appears in a few seconds, it will take a hundred steps to get there. Researchers at the Massachusetts Institute of Technology (MIT) have developed a new system to reduce the number of steps and achieve 30 times faster image display.

A new technique called Distribution Matching Distillation

As explained by our colleagues at Future, the creation of an image by an AI goes through several stages. For example, a tool like DALL-E retrieves an image from its database that is close to what the user requested. Then it is destroyed and a random noise field is created. Several other steps are required to eliminate random noise and create an image that matches the prompt.

Advertising, your content continues below

MIT researchers were able to reduce all these steps to just one using a technique called Distribution Matching Distillation (DMD). This method is divided into two components. The first is called regression loss; during this step, the images are organized based on their similarity during the training phase.

The second component is distribution matching loss. It allows us to ensure that the image reflects as much as possible what exists in our world.

Images are generated 30 times faster

In summary, the DMD technique allows the AI ​​to retrieve the maximum number of images that match the request from its database. The model then selects the models that best match what is found in the real world.

Images are generated more than 30 times fast, while reducing the risk of strange renderings. The required computing power is also decreasing.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *