I Tried Dall-E For The First Time, And It Was….
Ahh Dall-E…though I am a tad bit late to jump on the Dall-E hype train, its better late than never to try it out; and oh boy did I have an experience!
What is Dall-E and How Does it Work?
If you don’t already know, Dall-E is an ai art generator developed by OpenAi, that uses NLP and deep learning to create visual based on any given prompt.
Based on my limited understanding of the subject, the way Dall-E works is that it processes the prompt given using NLP and uses each word to take pictures from across the internet and mash it into a rough, low-res overview. Then, much like Stable Diffusion, it slowly processes that overview until it becomes a clear image; which is then outputted as the result.
Dall-E uses machine learning to constantly improve itself, so everytime you generate something, it learns from things like the key-words you used, the way you structured the prompt, the result, etc., so you will never get the exact same result twice.
Experimenting With Dall-E
Finally, the fun stuff! Pushing the limitations of the software was actually not that simple; Dall-E is really advanced, and only lacks in certain areas.
One of the areas the current build of Dall-E lacks in is facial features, specifically human faces. Take for example the prompt;
“the photo of an old man inside a heart-shaped locket, photo-realistic”
Which produced this Eldritch abomination:
Now, I understand what Dall-E is trying to create here, but by God does that face give me the heebie-jeebies.
Sometimes, this weirdly uncanny distortion of the human face lends itself really well to produce a beautiful composition. Especially for the prompt;
“a man trapped inside a marble, photo-realistic, intricately detailed”
This produced a really beautiful and oddly satisfying result, and may I add, in a very unique and modern art style. Truly a masterpiece.
Another feature which I found queer, to say the least, was the way Dall-E interprets different styles. Sometimes, you require something in a very specific style, for example, in the prompt;
“retro synthwave render of an explosion”
There is a very clear style; “retro synthwave”. Using that prompt, Dall-E created this absolutely wonderful image;
This image perfectly captures the desired style, “retro synthwave”. This is Dall-E working perfectly.
Sometimes, Dall-E decides to blank out, and creates weird amalgamations, for example;
“an extremely intricate, detailed nebula in the style of Spider Man into the Spider-Verse”
In the prompt, the style is specifically, “like Spider Man into the Spider-Verse”, which in my opinion is fairly recognisable. Unfortunately, Dall-E produced this weird image:
Instead of creating a nebula in the style of Spider Man, it has created a Spider-Man/nebula hybrid.
So far the majority of these examples have been bad, but most of time Dall-E truly delivers on its promise and creates some really astounding pieces of imagery. In that spirit, let us take a look at some of the absolute gems I created using Dall-E!
In conclusion; Dall-E is an amazing piece of software, and I highly suggest trying it out for yourselves. It is incredibly powerful and will only get better with time. I cannot wait to see what OpenAi comes up with next, but for now, this is THE END.