Lost in the Desert with DALL·E
An Art Exhibit of Ridiculous DALL·E Fails
DALL·E is Dalí.
My AI journey entered a psychedelic phase when I began experimenting with DALL·E. Letters melted. Words defied gravity as they floated in the sky.
If the images created by DALL·E seem inspired by the paintings of the Spanish surrealist artist Salvador Dalí — it is because they are. That’s why they named it DALL·E.*
DALL·E makes it easy to bend reality. Given the roiling, gobsmacking nature of news today, that isn’t necessarily unwelcome. We all could use an artificial vacation of our own making. OpenAI’s image generator transforms a description in natural language into photorealistic images.
You type: it renders.
With the introduction DALL·E 3, which is built natively on ChatGPT, users can collaborate with it and refine their prompts. OpenAi claims DALL·E 3 “significantly improves” on DALL·E 2.
Let you and me be the judge.
DALL·E 2
DALL·E 3
Misimpressionism
Based on the images, the primary difference between DALL·E 2 and DALL·E 3 is artistic style. DALL·E 3 ventures beyond impressionism’s visible brush strokes and surrealism’s images that defy the laws of physics. DALL·E 3’s style reflects a new movement in the art world: “misimpressionism,” a term coined by author and consulting professor Dr. Sam Savage. The Art of AI appears hyper-real and wildly false at the same time.
“I call this Misimpressionistic Art. For comparison, Impressionistic Art, exemplified by the works of the French nineteenth century artists, Monet and Renoir, is characterized by rendering the fundamental truth of a scene through the quality of the light, and other holistic features. But when you look closely, instead of fine detail, the brush marks are clearly visible without strong pattern definition. The Misimpressionistic Art of AI, on the other hand, uses nearly photographic detail to render fundamental falsehoods.
Dr. Sam Savage, Author of The Flaw of Averages, Consulting Professor of Management Science and Engineering at Stanford University
Viscerally, I find the image on the right disturbing. DALL·E 3’s over-amped fake realism sets up a buzzing discordancy when I view it. Its milk detonation comes across as aggressive. If DALL·E 3 were an artist, I might suggest he not take a psychedelic (LSD), an empathogen (MDMA), and a stimulant (methamphetamine) all at the same time.
It is simply too much, yet here we are.
DALL·E 3’s cookie isn’t wet from dunking. In fact, this is no dunk: the milk is exploding. Where, exactly, does the glass leave off and IEDed splash begin? And what’s up with the hyperactive psychedelic swirls in the background?
A pressured mania seems to be driving DALL·E 3’s cookie dunk tableau.
Yet some marketing maven at OpenAI believes this is art — significantly improved art compared to what came before. OpenAI asserts the image on the right is indubitably better than the impressionistic brushstrokes of DALL·E 2.
And yet, I prefer a kinder, gentler dunk.
In addition, OpenAI claims DALL·E 3 images offer greater resolution. (More pixels!) But more res isn’t always better.1 Additionally, users complain DALL·E 3 images decline in quality after extensive use. In fact, users have gone on to complain that DALL·E 4o images are very low resolution.
Then there’s that pesky text problem. DALL· E finds it virtually impossible to generate text in images, such as captions for cartoons or words on signs.
I discovered the issue when I asked DALL·E to generate an illustration for a recent blog post. I simply couldn’t grok how DALL·E could misspell the words I’d given it. My repeated attempts to get DALL·E to correct its mistakes resulted in images that were as flawed as they were surreal.
I hope the image sequence below will entertain and serve as a cautionary tale: first, read the directions.
"What we've got here is failure to communicate."
It all started innocently enough.
I asked ChatGPT’s DALL·E to create an illustration for AI and The Lying Game.
Create an image of a diner called "Citations" alongside a barren road in the desert with a flashing arrow-shaped neon sign that says "Eat Here" pointing at the restaurant.
What’s wrong with this picture?
The roadside diner is not beside the road, it is on it. Also, in the larger neon sign, “CITATIONS,” a blurred letter “A” shifted right up against the letter “T.”
I attempt to clarify my request:
Put the restaurant beside, not on top of the road. Have the Large Neon Sign above the restaurant say "Eat Here" and make the sign's shape more like an arrow. Remove the other "Eat Here" sign.
What’s wrong with this picture?
DALL·E named the diner “Cat Here” instead of “Citations,” as I’d requested. “EAT HERE” erroneously points away from the restaurant. A weird shadow from an invisible thing crosses the road.
I attempt to clarify:
Rotate the "Eat Here" sign so it is vertical, pointing at the building. Move the sign beneath it so it sits on top of the short side facing the viewer. Change the wording of that sign to say "Citations".
And this picture, what’s wrong? Almost everything.
(Are you frustrated yet?)
Let's try again from scratch: give me an image of a diner beside a roadway in the desert. Mounted on top of the building is a large neon sign shaped like an arrow, pointing down at the restaurant with the words "Eat Here". Another sign on the face of the building displays the word "GPT Citations"
They went thataway.
The arrow points to the left instead of pointing at the building,
I revised the prompt, attempting to clarify:
Rotate the arrow-shaped "EAT HERE" sign so it is vertical and points at the building. Remove the smaller sign "GPT Citations". Change the words on the rectangular "EAT HERE" sign to read "GPT CITATIONS".
Wow. Letters become airborne.
DALL·E generates unrequested skywriting and misspelled white signs.
Remove the floating GPT GPT and the two white signs to the right of the arrow. Make the GPT sign on the front of the diner much larger and have it say "GPT CITATIONS”
They went thataway the other way.
The arrow is not pointing at the building as I’d requested. More skywriting appears. A redundant sign sits underneath “EAT HERE” — only it says “E T HERE”
Let's try again from the beginning. Create an image with a diner beside a road in the desert with the name "GPT CITATIONS" on the building. A large neon sign in the shape of an arrow points at the building with the words "EAT HERE"
Still not pointing at the building.
Simply rotate the arrow sign so it is vertical and points at the building with just the words "EAT HERE". Instead of just "CITATIONS", make the name of the diner "GPT CITATIONS".
Arrow points thataway the other way again.
“Eat Here” goes missing.
Position the neon arrow so that it points at the roadside diner with the words "EAT HERE". Put a second sign on the face of the building with the restaurant's name "GPT CITATIONS".
“CPTS”? Honestly, what is that?
Also, “GPT Citations” makes an unwelcome appearance in the distance.
Have one neon sign in the shape of an arrow saying "EAT HERE." Have a second sign feature the name of the roadside diner, "GPT CITATIONS"
Um, where did the road go?
More sign redundancy.
The arrow sign needs to point at the restaurant -- it needs to be vertical pointing downward with the words "EAT HERE". A second sign is positioned on the front of the restaurant with the diner's name, "GPT CITATIONS".
So close but so far away.
“Here” is spelled “HRE”. The word “Citations” is spelled “CITTIONS”
Keep the design. Just correct the spelling. Make a correction to the spelling on the arrow so it reads "EAT HERE". Make a correction to the spelling of the restaurant name so it reads "GPT CITATIONS".
Let’s just add “HAE” for the hell of it. Got it.
I am out of ideas.
Try again
Actually, just add “HRE” for the hell of it.
Make a correction to spelling. On the neon arrow sign, change "HRE" to "HERE" so the sign reads "EAT HERE". Keep the design the same.
Bollocks.
“Citations” is misspelled again and again. The arrow needs to point at the restaurant.
Try again. There is an EAT HERE neon sign over the roadside diner. It is shaped like an arrow and pointed at the building. The restaurant name sign is "GPT CITATIONS" on the front of the building.
Dagnabbit.
Make an image of a roadside diner in the desert. Position an arrow-shaped neon sign so it points at the diner with the words "EAT HERE". Position a second sign with the name of the diner, "GPT CITATIONS," on the front of the building.
Poppycock.
try again please
Absolute codswallop.
To be more specific, I decided to highlight an error with an image for correction. I chose an image that required 1 spelling correction: change “HRE” to “HERE”
All I needed was an additional “E”.
After another 8 rounds and some 3 dozen prompts, I never generated an image that I could use.
I remain lost in the desert with DALL·E. And so far, according to OpenAI, there is no escape.
You might be tempted to instruct DALL·E to generate text in your image, by giving it instructions like "a blue sky with white clouds and the word hello in skywriting". However, this is not a reliable or effective way to create text. DALL·E is not currently designed to produce text, but to generate realistic and artistic images based on your keywords or phrases. While you can request text in your image descriptions, the results might be distorted, unclear, or not as expected, as it does not have a specific understanding of writing, labels or any other common text.
— Getting Started with DALL·E, OpenAI**
*I just figured out this morning that DALL·E is inspired by Salvador Dalí.
**I just figured out this morning why DALL·E has trouble generating text.
Pogue, David. 2013. “Why Hi-Res Isn’t Always Better.” Scientific American. April 1, 2013. https://www.scientificamerican.com/article/pogue-why-hi-res-isnt-always-better/.