KOALA Takes the Stage: New Open-Source AI Creates Images in a Flash


Researchers have developed a revolutionary AI tool capable of generating images in under two seconds, and best of all, it runs on affordable hardware. This breakthrough overcomes the limitations of previous models, which often required expensive and powerful computing systems.

The key to this achievement lies in "knowledge distillation," a technique that effectively compresses the knowledge of a larger model (Stable Diffusion XL) into a smaller, more efficient one (KOALA). 

This leaner version, with only 700 million parameters compared to the original's 2.56 billion, requires less processing power and memory, making it suitable for everyday computers and even low-cost GPUs.

The research team, from the Electronics and Telecommunication Research Institute (ETRI), has published their findings and made their work openly available for further exploration and development. They have even created multiple versions of their models, including KOALA for image generation and Ko-LLaVA for answering questions visually.

Testing demonstrates the speed advantage of KOALA. It generated an image based on a complex prompt ("a picture of an astronaut reading a book under the moon on Mars") in just 1.6 seconds, while comparable models like DALL-E 2 and 3 took significantly longer (12.3 and 13.7 seconds respectively).

With its impressive speed and accessibility, this new AI technology has the potential to revolutionize various fields, from image generation services and education to content creation and beyond. The researchers are actively exploring how to integrate this technology into existing applications, opening doors for exciting future possibilities.