In September 2023, OpenAI announced a new text-to-image model, Dall·E 3. I first got the opportunity to try Dall·E 3 when my company gave access to ChatGPT Pro, which includes the Dall·E 3 model. I am pretty excited because the images generated are far better than Dall·E 2 that I've explored some time ago.
So today, I created a simple python program and incorporated OpenAI's Image Generation API to redo the prompt from the previous Dall·E 2 experiment. Note that the Dall·E 2 images shown below were generated in November 2022.
Below is the image comparison of Dall·E 2 and Dall·E 3 side-by-side. Sorry for the inconsistent image sizes.
| Dall·E 2 | Dall·E 3 |
|---|
 |  |
| Totally unsatisfying, the picture is not realistic at all. | The picture is much more realistic and nicely composed. Also, cute. |
2. A digital art of a lighthouse with the Andromeda galaxy in the background
| Dall·E 2 | Dall·E 3 |
|---|
 |  |
| Good, just a bit cartoonish | Satisfied. Nicely composed. But it looks like Earth has no atmosphere at all (The galaxy should look faded, but I assume maybe it uses a long focal length setting) |
| Dall·E 2 | Dall·E 3 |
|---|
 |  |
| Ok good enough | Better, have more life |
| Dall·E 2 | Dall·E 3 |
|---|
 |  |
| Looks stressful, apocalypse-like scenario (maybe because this image was generated just after the spike of the Covid-19 era) | Future-modern life like depicted in most of my primary school textbooks |
5. Profile picture with a man and cat in a tech environment
| Dall·E 2 | Dall·E 3 |
|---|
 |  |
| Good enough, but why is the cat sad though | Not composed nicely. The cat looks more realistic than the human. |
Overall, I think Dall·E 3 is now able to generate results that are richer in context and eye-pleasing. For those who want to see the python script used to generate the images, please refer to the GitHub repository openai-dalle2 & openai-dall-e3.
