South Korean scientists from the Electronics and Telecommunications Research Institute (ETRI) have developed artificial intelligence (AI) technology to create images almost instantly. According to the creators, the model works five times faster than existing analogues. The article was published on the official site Web site National Science and Technology Council (NST).
Experts presented three models based on the KOALA neural network, as well as two models of the interactive visual language KoLLaVA, which can answer user questions using images and videos.
Using a method called data distillation, ETRI was able to make KOALA significantly more compact compared to existing AI graph generators.
Thanks to this, the model can run on relatively inexpensive graphics processors with eight gigabytes of memory. The model creates an image with high detail and resolution in just 1.6 seconds. By comparison, OpenAI’s popular DALL-E 2 neural network takes 12.3 seconds to complete the same task.
ETRI has also launched a website where users can directly compare and test a total of nine models, including two publicly available stable diffusion (AI rendering) models: BK-SDM, Karlo, DALL-E 2, DALL-E 3 and three KOALA models.
In the future, the research team expects high demand for Korean cross-modal (using different types of data) models that integrate visual intelligence technology into open-source artificial intelligence.
Developers before was created A neural network that generates background sounds for videos.