You are here
First Thoughts on DALL-E v3
Posted by Jim Craner on October 31, 2023
We're very cautious about using AI image generators here - they can be incredibly fun and powerful, but there are some serious questions about the ethics of training image-based AI on public images without the copyright holders' consent. We thoroughly cover the copyright and intellectual property issues, as well as the economic displacement of humans, in all of our AI workshops.
The newest image generator from OpenAI, DALL-E 3, is designed to alleviate some of those intellectual property concerns. From the linked page:
DALL·E 3 is designed to decline requests that ask for an image in the style of a living artist. Creators can now also opt their images out from training of our future image generation models.
OpenAI seems to have refined this feature over the past couple of weeks. When I asked for a painting "in the style of Matisse" on October 24th, I was told that DALL-E 3 "can't reference Matisse directly as he's a relatively recent artist." For the record, Matisse died in 1954 so there's one data point for what DALL-E 3 considers "recent" in this context. I'm not sophisticated enough to know many visual artists who have died in the past few decades so I didn't test with other prompts.
Note that it did offer to write a new prompt based on Matisse's style!
Since then, it appears that they have refined the behavior when prompting for recent artists. DALL-E 3 no longer refuses to fulfill the request. Instead, it immediately rewrites the prompt based on the artist's style and creates the images. Here's a similar request initiated on October 31; note that the actual prompts generated (numbered "1" and "2" still reference "in the style of Henri Matisse" as in the test from earlier.
This feature seems inconsistent to me. It agreed at one point to generate an image transformation in the style of Basquiat (d. 1988), but errored out. When I asked for that image again, I was told:
I'm sorry, but I can't create images in the style of Jean-Michel Basquiat or directly reference his work since he was active within the last 100 years.
For the record, the latest work I could find from Matisse was 1953, which is definitely within the last 100 years.
Ripping off the style of living artists is only one of the intellectual property concerns that copyright holders have with image generators like DALL-E 3. For instance, DALL-E3 should officially reject requests to generate images containing copyrighted characters. But just a few experiments have mine have yielded images with copyrighted characters identified by name and clearly resembling their counterparts.
Results for "Bob Belcher in a suit of armor" -- successfully generated with DALL-E 3.
Now that we've gotten the intellectual property concerns out of the way, I will say that DALL-E 3 is absolutely the coolest, best image generating model I've ever seen, hands down, by far. By accessing the image model via the ChatGPT interface, we get ChatGPT's "take" or "interpretation" of our prompt. This can result in some weird and inconsistent images but also some very creative compositions that you might not have originally envisioned. Here are just a few from our early experiments:
DALL-E 3 is available to paid users of the ChatGPT Plus application.