Quick Links
Generative AI art has come a long way since last year. But exactly how far has it come? I was wondering the same thing, so I checked out my old DALL-E 2 prompts and their resulting images, and then compared them to what DALL-E generates today. Well, the results are pretty stark.
1Map of a Rainforest Land
Last year, I tasked DALL-E 2 with creating a rainforest map to serve as inspiration for a novel I was writing. The outcome it produced was basic, resembling a simple doodle I might have crafted in Paint in my childhood.
While the text in the newer AI image is still nonsensical, the sheer difference in quality and depth between the two is quite overwhelming.
Instead of the flat paper style of the DALL-E 2, the new DALL-E (which you may access via aChatGPT Plus subscription) preferred to give it a 3D-like spin, reminiscent of illustrated encyclopedias. The words in the newer image are borderline readable, and this was definitely a better attempt at representing insects, animals, trees, and fauna.
DALL-E 3 is now known simply as DALL-E and is only available through OpenAI’s ChatGPT.
2Pixel Rose on a Cloud
DALL-E can be useful for creating wallpapers. So, while brainstorming ideas for a wallpaper, I asked DALL-E 2 to create pixel art of a red rose growing from a cloud. I wanted the effect of the cloud acting like soil for a rose. It managed to produce the elements I wanted, but it didn’t interweave them like I had hoped—even after repeating the prompt multiple times.
However, the new DALL-E was up to the task, bringing characteristic depth into its interpretation of the prompt. Not only did it actually embed the rose into the cloud, but it added many more details to the flower, fleshing it out to be more than just a basic icon. But I noticed it added elements I didn’t ask for, like the floating pixel hearts hovering above the rose.
3Black American Female President
DALL-E 2 always had mixed success with doing human faces; sometimes, it would look scrambled and off, and other times, it would look so lifelike that you’d wonder if it wasn’t someone’s real photo. My prompt for a female Black American president was one of those that struck me how believable it turned out.
When it was DALL-E 3’s turn, though, it came out in that plastic-like 3D animation style it’s prone to. That said, it did a better job of creating a context for our generated president by placing the subject on a podium and putting White House columns in the background. However, I can’t help but give the realism prize to the older version of the AI.
4Tiny Man in Van Gogh’s Style
I wanted to see how well DALL-E 2 could imitate another artist’s style with an unfamiliar, original subject matter. As you’re able to see, it got the color right; it seemed to have drawn the characteristic yellow from the many wheat-field-like paintings Vincent Van Gogh had created.
However, the style is far too rough and disorganized, and an onlooker might struggle to realize what artist inspired the piece. Also, it didn’t do a good job of representing a tiny man fighting ants in the field; instead, it looks like a farmer working a field.
On the other hand, DALL-E 3’s image was crisp and clear, sporting none of the graininess that DALL-E 2 prefers. The subject matter is undoubtedly a small man fighting ants on a field.
However, apart from the Starry Night sky in the painting, I would be hard-pressed to call this Van Gogh’s style. DALL-E 2’s imperfections seem to make it a better candidate for creating paintings. DALL-E 3’s painting comes out looking too polished, leaning too much into the digital asset aesthetic.
In any case, DALL-E 2 is no longer available, and all its features and capabilities are the foundation of DALL-E 3. Whether DALL-E 3 is completely better may be debatable, but it’s certainly better at making more comprehensive, higher-quality images.
