Researchers from South Korea’s Electronics and Telecommunications Research Institute (ETRI) have unveiled an AI technology capable of generating images in near real time. The team reports that their model operates five times faster than current equivalents, delivering highly detailed visuals at a speed that stands out in the field. The development was announced on the National Science and Technology Council’s official platform, highlighting the practical implications for fast, resource-efficient AI image generation.
During demonstrations, three variants based on the KOALA neural network and two interactive visual-language systems built with KoLLaVA were shown. These tools can answer user questions by analyzing accompanying images or video content, enabling more dynamic human-AI interactions and expanding the range of multimedia prompts that can be understood by the AI.
A key technique behind the breakthrough is data distillation, a process that makes the KOALA models more compact without sacrificing performance. This compression allows the AI to run on more modest hardware, including graphics processing units with memory as low as eight gigabytes, broadening accessibility for developers and researchers who do not rely on high-end infrastructure.
In practical terms, the system can render a high-detail image with impressive resolution in about 1.6 seconds. By contrast, a widely used contemporary model from another major player in the field typically requires around 12 seconds to reach a comparable result. The efficiency gap underscores potential savings in compute time and energy, making fast AI image generation more feasible for real-time applications and interactive workflows.
ETRI has also launched a public testing portal where users can compare and evaluate nine different models side by side. Included are two openly available Stable Diffusion variants, along with popular models such as BK-SDM, Karlo, DALL-E 2, DALL-E 3, and three KOALA configurations. The portal serves as a practical sandbox, inviting researchers and practitioners to assess capabilities, limitations, and how each model handles complex prompts and cross-modal inputs.
Looking ahead, the research team anticipates strong demand for cross-modal Korean AI systems that integrate visual understanding with other data types and open-source software. The aim is to advance open architectures that enable robust visual intelligence across a range of domains, from creative design to data visualization and educational tools.
Earlier efforts by the same group included developing a neural network capable of generating background soundscapes for video content, signaling a broader push toward integrated, multimodal AI that can handle sensory data in a cohesive way. The ongoing work reflects a clear trajectory toward systems that blend sight, sound, and language to support richer, more intuitive human-computer interactions.