An artificial intelligence can guess what a plate of food will look like just by reading the recipe. Many of the images it generates look realistic, but others look barely edible.
The AI was trained on 52,000 pairs of recipes and images of the resulting plate of food. One part of the system learned to create images by reading the recipes, while a second attempted to distinguish the generated images from real ones. By repeatedly trying to make images indistinguishable from real ones, the AI was able to improve the results.
Although some of the images looked appetising, many look like a close-up of gooey ingredients. A few of the images look like food you might cook at home but would never share on Instagram.
Advertisement
Part of the problem is that the images used to train the networks were low quality. 鈥淭his is reflected by lots of blurred images with bad lighting conditions, 鈥榩orridge-like images鈥�, and the fact that the images are not square shaped which makes it difficult to train the models,鈥� wrote the team from Tel-Aviv University in their paper.
听
That could account for the fact that both models had an easier time producing foods with less structure like soup, rice or pasta, and struggled with things like hamburgers or chicken.
The team asked 30 people to score the quality of the images. On average, people rated the closeness of the images to the original recipe as 2.9 out of 5, and the resemblance to real food as 3.7 out of 5.
AI has previously been used to generate images from short sentences, but recipes contain far more complex text, including the ingredients list and several multi-step instructions.
The idea came when team member Ori Bar El asked his grandmother for the recipe for her 鈥渓egendary鈥� fish cutlets with tomato sauce. She couldn鈥檛 remember the exact recipe, so he thought about creating a system that would output a recipe from an image, but decided it would be too hard. So instead created a system that would do the reverse.
Reference:听arXiv,听
