On April 22nd, Marina Pantcheva and I gave a workshop at a conference called GALA 2024. GALA is short for Globalization and Localization Association, and is the largest industry association that brings together firms and professionals who deal with translation and localization. Now the translation industry is affected by the ascent of generative AI more than others, and there is significant apprehension to what the future might hold for the professionals and companies in this field.
In this environment, a workshop on how to master AI prompts and how to understand generative AI better can have a lot of value — and that was exactly what we set out to do. This post is about the second half of the workshop where we prompted images with DALL-E 3 to understand better how a multimodal generative model responds to various prompts. (In the first half of the workshop, Marina showed a very powerful series of prompts that help language professionals run quality checks on translated content.)
In the second half, the participants chose one of three photos. The assignment was to write a prompt that most closely replicates the original photo. For copyright reasons, I cannot replicate the photo here, but here is the link to the original Instagram post. The photo shows a tram pulling in a station in the rain, and a person with an umbrella is waiting for it. The rain is so heavy that the view is blurred.
We have collected the prompts from the audience through Sli.do and we were able to run some of them through DALL-E and look at the results together.
After the workshop, I fed all the prompts to DALL-E. What follows is all of these prompts and their results. You are welcome to compare – but, like always, you cannot expect DALL-E to replicate the exact same image when you give it the same prompt weeks later. Also, don’t expect a detailed analysis: there is no space for that in this post. I did not change the spelling of the prompts, either.
“Black and white image
(man with umbrella)
(tram on rails)
(heavy rain)
add vertical grain”
“Create a BW image of tram with a tram in rain, people with umbrellas, late in the evening. Stylize the image to look as if it’s drawn with a pencil.”
“Create an artistic black and white photo depicting Prague in the rain with a tram at a tram stop and a man with an umbrella.”
(The second tram is more like a hovercraft, isn’t it? No wheels.)
“Create a grainy b&w image of a 1920s tram in the rain with person waiting to get on”
“Photorealistic image in black and white of a long road with rain, a tram on the road and a person with an umbrella standing on the curbside”
“A photorealistic black and white portrait image depicting a train in a rainy street”
“Produce a picture of a street on a rainy day, some people walking around with umbrellas and a tram. It should be black and white.”
“A black and white photorealistic picture of a street in London with a train passing and Victorian lamp post, and on the left, there is a man standing with an umbrella in a very heavy rain.”
“Create a black and white realistic photo depicting a heavy rainy day in London, with people holding holding umbrellas and a tram passing by.”
Finally, here are the images where the prompt deviated from the original theme. This happened when the prompter forgot something (for example, “black-and-white”) or when they decided to follow their imagination, which is OK and makes the exercise interesting. I am not including the prompts with these.
thirdtower.ai 