How will an algorithm from the 21st century recreate dishes cooked in traditional Chinese style from a recipe book written in the Qing Dynasty? “The New Way of Eating” is a selection of reinterpreted recipes taken from “The Way of Eating”, a book written by poet and scholar Yuan Mei.
The initial version of “The New Way of Eating” contains of 7 recreated recipes and a cooking video showcasing the instructions and results for making one of the dishes, “Hibiscus Pork”.
Going with the theme of altering something well known by people into something unknown and disturbing, I started to research famous cook books and recipe collections, and finally settled with “The Way of Eating” (or “Recipes from the Garden of Contentment”). Originally published in 1792, it is a book that contains instructions and critiques on Chinese cuisine. I decided to create mockup cookbook spreads for a selection of recipes.
The prototype is divided into two parts: creating directions and ingredients (text generation), and creating images for the dishes (image generation from text). I selected 7 dishes from the book as a starting point, and used GPT3’s pre-trained model and preset “recipe generator” for text generation.
For image generation, I followed this paper “Zero-Shot Text-to-Image Generation” and its PyTorch implementation, then modified it to fit this project.
To first train the VAE, I scraped 5712 food images from Food Network as input data. Then to train DALL-E, I used the titles of the food images, and generated 9 256 x 256 images for each of the selected dishes. For the final layout I combined the images and generated instructions for the spreads, and followed one of the recipe, “Hibiscus Pork” and tried making the dish myself.
Other Example Spreads