A South Korean scientific team recently developed a KOALA The new AI image generation model significantly reduces the hardware requirements.And it can generate high-quality images within 2 seconds.
The key to this model is the use of a new technology called "knowledge distillation", which greatly compresses the size of the open source image generation tool Stable Diffusion XL.
Stable Diffusion XL currently has a total of 2.56 billion parameters, and the Korean scientific team used "knowledge distillation" technology to reduce the parameters to 700 million.
Therefore, the KOALA model does not require high-end graphics processors and complex equipment to run smoothly. It only needs 8GB of memory to generate images, and the generation time is shortened to within 2 seconds.
Essentially, knowledge distillation filters information from a large model into a smaller model without sacrificing quality and performance. This allows the smaller model to generate high-quality images faster.
According to the team’s test results, with the same prompt of “a picture of an astronaut reading a book under the moon on Mars”, the KOALA model takes 1.6 seconds to generate, while OpenAI’s DALL-E 3 model takes 13.7 seconds and the DALL-E 2 model takes 12.3 seconds.