Artificial intelligence research group OpenAI has revealed an impressive new AI that can create photos and artwork based on text descriptions.
Back in February of 2019, artificial intelligence research group OpenAI announced that it had created an AI so dangerous that researchers refused to release the fully trained model. Named GPT-2, the AI was trained on millions of pages of text from the internet and given a simple goal: Given a sample of text, predict the next word.
GPT-2 was considered dangerous because it was able to generate believable articles and messages on any topic, and researchers feared that it could be used to create automated misinformation campaigns or harassing messages. The model was eventually released and superceded by the even more impressive GPT-3 model, which was trained on much more data.
RELATED: OpenAI creates an AI so dangerous it's refusing to release it
OpenAI has now announced another impressive breakthrough with the release of DALL·E, a neural network capable of generating photos and artwork from text prompts. The AI is a version of GPT-3 that's been trained on data from millions of images and their associated text descriptions.
DALL·E is able to take a short text prompt such as "an armchair in the shape of an avocado" and will then generate a series of images corresponding to that input. This works equally well to generate realistic photographic results and illustrations, though the results are still prone to some visual mistakes.
RELATED: Is artificial intelligence the future of journalism?
While AI has been used to generate images before, this is the first time that a deep-learning system has used a text prompt to generate imaginary photos and drawings in such accurate detail. Researchers were also able to prompt the model by supplying a photo and telling the AI to modify it, such as putting text on an item or changing its colour or style.
The model can also be prompted to create more specific images by giving it the first few lines of an image along with the text prompt, making it imagine the rest. OpenAI has released a detailed article on its findings, along with thousands of images and interactive prompts you can experiment with on their website.
Source: Open AI
