new Artificial intelligence (AI) The tool can generate images within 2 seconds and does not require expensive hardware to run.
Scientists in South Korea used a special technique called knowledge distillation to compress the size of an open-source (or publicly available) image generation model known as Stable Diffusion XL. This model has 2.56 billion parameters, or variables that the AI uses to learn during training.
The smallest version of the new model, known as “KOALA,” has just 700 million parameters. That means it's lean enough to run fast without requiring expensive, energy-intensive hardware.
Related: AI chatbots need to become even better at remembering things. Have scientists solved their terrible memory problem?
The method they used, knowledge distillation, transfers knowledge from large models to small models, ideally without compromising performance. The advantage of a smaller model is that it takes less time to perform calculations and generate answers.
The tool can run on low-cost graphics processing units (GPUs) and requires approximately 8 GB of RAM to process requests. Larger models, on the other hand, require high-end industrial GPUs.
The team published their research results as a paper on December 7, 2023 in a preprint database. arXiv. They have also made their work available through an open source AI repository. hug face.
The Electronics and Telecommunications Research Institute (ETRI), the agency developing the new model, has released five versions, including three versions of the “KOALA” image generator, which generates images based on text input, and two versions of the “Ko-LLaVA”. Created a version. Answer text-based questions with images and videos.
When tested, KOALA produced an image in 1.6 seconds based on the prompt: “Photo of an astronaut reading a book under the Mars moon.” According to one source, his DALL-E 2 in OpenAI generated an image based on the same prompt in his 12.3 seconds, and his DALL-E 3 in his 13.7 seconds. statement.
The scientists now plan to integrate the developed technology into existing image generation services, educational services, content creation and other business areas.