SEOUL, April 20 (Korea Bizwire) — Kakao Brain, an artificial intelligence (AI) unit of South Korean tech giants Kakao Corp., said Tuesday it had unveiled an AI model designed to create images from text descriptions through the open-source community GitHub.
The RQ-Transformer, consisting of 3.9 billion parameters, is a text-to-image AI model that learned 30 million pairs of texts and images. It can reduce calculation costs and improve the speed and quality of the creation of images.
The new AI, an upgraded version of the ultra-large multimodal AI “minDALL-E,” features a model size that is three times larger, as well as a dataset that has doubled in size, and an image creation speed that is twice as fast.
It’s an image creation model that learned how to create three-dimensional code map images through sequential prediction.
Compared to existing technologies, it can reduce the loss of image compression, thereby being able to create high-quality images in a low-resolution code map.
This model, accordingly, can reduce calculation costs and improve image creation speed compared to existing image creation models.
After learning through large-scale datasets, the RQ-Transformer can understand the composition of text it has never seen before and create images that match well with the text.
Kevin Lee (firstname.lastname@example.org)