The AI Recipe Book
Text generation models such as GPT-3 have become increasingly powerful, and have a large variety of use cases.
I've played around with them for a long time, doing several experiments in text generation, from writing Instagram comments to creating restaurant reviews, to simulating the writing style of a specific author.
With the release of DALL-E 2, which lets you generate images from text prompts, I had the idea of combining AI text and image generation to create a full recipe book. I can use GPT-3 to generate recipes and DALL-E to generate the images for the recipes.
- Classic dishes
- Fusion dishes
- Celebrity chef's Signature dishes with a twist
1. Generating the recipes
I've used GPT-3 to generate the recipes, however, a lot of trial and error is needed for it to generate good recipes!
The model already has a very thorough knowledge of cooking and foods in general, and I have tried to push it to be more and more creative.
First, by generating just classic recipes, then by combining different cuisines into original fusion dishes, and lastly to create recipes from famous chefs, but adding a twist to them.
Not only is it possible to generate original recipes with original flavor combinations, but it's also possible to simulate the writing styles of different chefs.
2. Generating the images
Generating food photography has been the main focus of my experiments with DALL-E.
I've been doing food photography for a while, so I have a clear idea of what a good food photo should look like, and DALL-E did not disappoint.
The biggest challenge, however, was to create realistic-looking photos.
One of the main issues was that the output image would look somewhat like a painting, or that some details or objects would not make sense.
A lot of trial and error was needed but I got to a point where the photos look almost indistinguishable from reality. I've actually been posting a few on my food Instagram account and nobody noticed.
3. Book Structure
I've decided to structure the book into 3 sections, aimed at pushing different capabilities of the AI models.
4. Classic dishes
This section is meant to be the first introductory test for the model.
It consists of several classic recipes for a few different cuisines.
This will test the general food knowledge of the model, and it's the first test for the image generation capabilities of DALL-E.
5. Fusion dishes
This section tests the creativity of the model.
It needs to be able to first generate original flavor combinations from different cuisines, and second, incorporate them into a recipe.
It's also a great example of image generation, where the model needs to create completely new photos of original dishes.
6. Celebrity chef's Signature dishes with a twist
This last section is slightly most subtle because it needs to incorporate small changes to a famous recipe, by still maintaining the style of the chef who created the original recipe.
No comments.