Ryan Mercer's thoughts, mostly random musings, spanning form 2001 to present. Freemason, geek, nutter, Whovian, 8-bit Atari enthusiast, SciFi fan.
I had the image of David the gnome riding a fox in my head and thought I’d task DALL-E 2 with something along those lines so I gave it the query “A photograph of a gnome riding a fox by a creek” and, well, what we got were essentially garden gnomes with scaled down foxes, still cool though!
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
A co-moderator of a discord I am on has been in Germany on a work trip, he’s been sharing with us all the gummies he has been eating and how he has 22lbs of them he is bringing back to the United States when he leaves tomorrow. In the discord chat someone joked that he’d eat them all before his submarine docked so I gave DALL-E 2 the prompt “A German submarine made out of gummies” and some look like artificial baits.
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Recently my wife and I visited Nauvoo, IL and stopped by the Nauvoo Illinois Temple and I took this photo:
I took this photo with my iPhone SE 2020
One of the features DALL-E 2 has is that you can upload a photo and either edit it, or produce variations. I opted to produce variations to see what it would do, and how well it could edit a real building - would it mess up the corners and windows like it does when it generates a building purely from your text input. Here are the results (spoilers, it still messes up architecture):
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
Created with DALL-E 2
So it would seem that generally DALL-E 2 is just not good at generating buildings, and possibly even vegetation as it has done some weird things with trees in some of these.
I wanted to see how DALL-E 2 handled food requests, I gave it the following query “A 7 patty cheeseburger surrounded by thick cut fries with a glass of soda” and thought it did alright. It even gave me the option of both regular and cheese fries in one image. Some of the glasses look a bit odd, that’s blatantly a can of Coke, and I don’t know what is going on with the bowl of cheese or butter in the one image but not bad. I will say it didn’t take the number “7” into account at all, although it was a very wordy query. I imagine with some refinement of the input you could get a much better image but not bad for my first cheeseburger query!
All of the below images were generated by DALL-E 2 from a text input.
Patty looks real and possibly like ribeye, fries look real, the glass is peculiar, the bun is almost perfect but the eggwash reflecting the light in the top right is ever so slightly off. Hands down this is the most realistic-looking burger of the lot.
Fries look real, that’s blatantly a coke can, the glass isn’t very good, I’m not sure what’s going on with the paper in the top right by the coke can, the bread and patty look fake but the plasticized cheese looks very real compared to cafeteria food as does the visually unappealing lettuce.
This table is beautiful, something weird going on with the chair, the fries look like they’ve accidentally merged with some sort of chicken or fish breading in spots, the bun looks more like a super chonky pancake, and the glass is perfect except for the weird ice anomaly at the back of the roughly 12 o’clock position where it looks like the ice physically goes through the glass, I’m not sure what’s with the slabs of butter or cheese in the mirraging metal dish though, and the burger patty reminds me of McDonald’s ‘steak’ in their old steak, egg, and cheese bagel.
This glass looks good aside from appearing like it’s been photoshopped in from another image. The cheese fries look good but overexposed, the regular fries have all sorts of weird things going on, the bun here looks very blah and the edge of the patty is just odd.
Someone smashed the top of the bun but it looks real, the bottom of the bun is a bit less real, the plasticized cheese looks real and the patty just looks like your generic cafeteria super processed patty. The fries, as with other examples, are just slightly strange looking in places and metal tray they are on looks good except for that “11” bar roughly in the middle. This glass doesn’t appear symmetrical but it’s one of the better ones of the lot, I would 100% hate that ice in real life though.
Bun, cheese, fries, and bowl all look fine. The glass again though has issues with the rim although the rest of it is probably the best of the lot and the ice looks decent especially given how opaque the drink appears to be. The patty though, the patty just looks wrong.
The other day Sam Altman’s tweet about Dairy Queen’s Blizzards got me curious (and hungry, grrrrrr, thanks a lot Sam!)
dairy queen blizzard is by far the best fast food
— Sam Altman (@sama) June 16, 2022
so I fed (pun not intended) DALL-E 2 the query “An artificial intelligence enjoying an oreo dairy queen blizzard”. This one was my favorite (and a rather interesting way to enjoy ice cream):
It generated some other nice results as well, just not as fun:
These are the best “eyes” I’ve seen DALL-E 2 generate to date as far as imitating photographed objects is concerned.
See how it does weird stuff with the left eye in this one as compared to the one immediately above?
Again, the eyes are a bit odd here, notice what looks like clumped clay/paint on both?
I found it interesting that it took “oreo” and created all of the entities with the white and black, it really seems to like blue too for this query, perhaps from the blue ‘swoosh’ in the Dairy Queen logo?
The Church of Jesus Christ of Latter-day Saints issues lots of artist renderings of forthcoming temples, has lots of images of their physical temples available online, and a lot of art created representing these buildings also gets shared online. With that in mind, I gave DALL-E 2 the prompt “The Church of Jesus Christ of Latter-day Saints, digital art” and was not disappointed
Created with OpenAI DALL-E 2
Created with OpenAI DALL-E 2
Created with OpenAI DALL-E 2
Created with OpenAI DALL-E 2
Created with OpenAI DALL-E 2
Created with OpenAI DALL-E 2
These are impressive, while it messes up lines and proportions a bit, it nails the actual style used for the real-world artist renderings of future buildings. For comparison, here are a few real examples of future buildings:
The DALL-E 2 generated ones just seem like rough drafts before a final version was put together, quite impressive!
I had hoped that the infant would be a grey alien, or something else, when I gave DALL-E 2 this prompt but, I’m still quite happy with the outcome.
Made with OpenAI DALL-E 2
It genuinely just blows my mind that software generated this in less than 10 seconds from a text prompt. I’m quite happy with the fact it took “carnation” and made some very strange planet in the top left. The random assortment of planets with orbits that make no sense whatsoever, combined with the carnation planet, even make it feel like something from the Renaissance or even Baroque movements.
As someone that loves vintage gaming consoles, vintage computers (the Atari 800XL is my machine of choice from that era, I’ve owned about a dozen over the years and have since downsized to just 2), and vintage monitors I decided to try this query. DALL-E 2 created these with the prompt “A photograph of a robot playing vintage gaming consoles”
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
I’ve noticed that DALL-E 2 seems to struggle with controls, in most of these the buttons, switches, and knobs lack symmetry and refinement. Often looking like a chunk of the everlasting gobstopper in the Gene WIlder film Willy Wonka & the Chocolate Factory. I had previously noticed that it also struggles with lines on devices and buildings… it seems to understand what an edge/corner/window is but doesn’t understand quite how they should look.
This is by far the simplest DALL-E 2 query I have input to date, simply “Neon Theotokos”
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
Made with OpenAI DALL-E 2
I honestly had no idea what to expect given I tried several other religious art queries prior to this with less than stellar results (so much so I didn’t even bother to save them) but I love ALL of these variations DALL-E 2 generated. This is such a neat tool.
