After the amount of criticism and especially memes that Uruguay’s last alternative shirt received, the idea arose that variants could be made. For this I relied on three tools for generating images from text DALLE, Crayon (former DALLE mini), midjourney, NightCafe and dream studio, from where some more designs emerged, from the already official shirt.
Before showing the results obtained by each one, a brief introduction of how text-to-image models work.
Although the architecture and the information with which they are trained vary, these models work with a quite similar logic, roughly speaking, we will say that they have a language model and an image model.
What the language model will do is assign a code (token) to each word, this process is known as encoding, so the system will understand what we write. Both DALLE mini and DALLE 2 use a system called CLIP, this system not only keeps the categories with the list of predefined categories but is capable of identifying new categories without human participation. Another advantage it has is that it generates a dictionary where it associates a word with an image.
The architecture of unCLIP. (Image source: Ramesh et al. 2022])
On the other side we have the image model. In its Blog, Daniel Fein describes this process as building a Rubik’s cube, since after encoding and subsequent decoding, what remains is a cloud of pixels, from there the model seeks to reconstruct the image towards what should be the final result.
The purpose of these models is not just to make funny images, but to seek to have a commercial application. The main use that has been given to them is to generate product prototypes, participate in the creative process or inputs for marketing, such as the cover of Cosmopolitan created by DALL-E2.
After this brief presentation of how the model works, now, the important thing, the new Uruguay alternative shirt.
All the results obtained were placing “Uruguay football away shirt”, “Uruguay football away white shirt” or “Uruguay football away red shirt”.
Here we show you the results obtained:
Craiyon or DALLE-mini is characterized by generating images that are not so exact or detailed, but very creative.
DALL-E achieves a more real and precise result
Midjourney has a more artistic look by default, which for me achieves the best results:
And finally, because we know that there are some defenders of the red alternative
Night Cafe and Dream Studio are based on a model called Stable Diffusion that was recently released.
Dream Studio with Stable Diffusion V1.5
Which tool do you like the most to use? Do you think the results were better than the official Uruguay alternate jersey?