A Recipe of ChatGPT, Stable Diffusion, and Canva Delivered a Cookbook in 1 Hour – Generative AI in Practice
Synthedia and Voicebot are conducting in-depth research on generative AI-based synthetic media solutions for text and image creation. Some of the findings will be revealed publicly for the first time on December 14th at the Synthedia online conference. Our research team looked at more than ten AI models in both the text-to-text and text-to-image categories to provide a comparative analysis of the solutions. Not all AI models produce the same output, even those based on the same underlying solutions.
Synthetic Media Demo with Generative AI
In addition to the data, we wanted to show how these tools could be used in concert to create a product you might see on the web or in a bookstore. A cookbook was an obvious choice because you need a significant amount of text alongside stunning images. This was a test of both quality and speed. We reserved just 30 minutes for the cookbook planning along with the text and image generation. There were then 30 minutes for creating the digital book. What is a better way for three colleagues to spend a Friday afternoon? One person even likened it to a cross between applied research and team building.
The text and images generated by ChatGPT and Stable Diffusion were published as-is, with no edits other than formatting the font size and style. We simply copied and pasted them into a Canva document. You can see all 12 pages by scrolling through the document above. The first three pages provide some context on the project. We then present three recipes–an entreé, a side dish, and a dessert. Each recipe includes a backstory in a style popular in modern cookbooks, the ingredients list, and preparation instructions.
What We Learned
The research team is already wading deeply into the generative AI waters, and we have experience creating reports in the digital book format. That certainly helped us move more quickly through the process. However, if you follow a similar process, you should be able to produce something comparable even if it takes more time. We also believe our lack of experience with recipe and cookbook production offsets those other skills somewhat.
Three findings rose to the top from this project:
- 30 minutes was sufficient time to decide on the type of recipes we wanted, define very brief backstories, and then generate the text and images.
- If we had more time in this section, we would have prioritized generating more image variants. After that, we likely would have done some editing on the backstories. Spending an hour on this section definitely would have resulted in a more refined output from a content standpoint.
- 30 minutes is not nearly enough time to do a proper layout and design, even using a tool such as Canva. Document design and layout are iterative processes. If you have more time for iteration, you will likely deliver a higher-quality output.
The goal was not speed. We wanted to see what type of quality you could deliver with three people working on a cookbook together for 1 hour. For the content generation, 10 minutes was estimated to create the content for each recipe, and that was just enough time, though it was rushed.
Learn More at the Synthedia 2 Online Conference
Synthedia is a daily newsletter on topics related to synthetic media, and it is also the name of a free online conference series. Synthedia 2, the event, will be held on Wednesday, December 14th. The creators of the 1-hour generative AI cookbook will share their experience with the process and answer questions. They will also reveal some of the results of the upcoming research reports on generative AI text-to-text and text-to-image models.
Other speakers include the founder of DeFiance media and the editor of Heise Online. They will speak about using virtual humans for video news and voice clones for podcasts, respectively. Altered AI will discuss scalable synthetic voice prototyping and production. Scott Stevenson will demonstrate how Rally Law’s implementation of GPT-3 will write and review contracts for lawyers. Michael Pierre Price will discuss his experience generating 1,001 images on Midjourney for a charity auction. AI analyst Bakz T. Future will review what is going on at OpenAI and the news around Stable Diffusion.
- Day: Wednesday, December 14th
- Time: 12:00n – 3:00p EST
- Location: Online Zoom event
I hope to see you at Synthedia 2 and reading the daily newsletter.
Follow @bretkinsella Follow @voicebotai
‘AI Howie Mandel’ by DeepBrain AI Joins Growing Cast of Synthetic Celebrities