DALL-E 3: What you can and can’t do with the image generator

DALL-E 3 is the third iteration of image-generating AI from OpenAI, the creator of ChatGPT. This is dedicated to the generation of images, unlike GPT which is oriented towards textual outputs. Like every AI released this past year, it opens up a world of possibilities, but not without certain limits…

Midjourney could generate photorealistic images, but DALL-E 2 lagged in this area, producing artificial images that were easily identifiable. For the release of the third version of DALL-E, a research article explains how it works. We reviewed this article to determine the limitations and areas of expertise of this AI.

Limitations of DALL-E 3

DALL-E 3 represents a significant advance in image generation, making it possible to produce images consistent with the generated texts. Until then, image generators had difficulty producing text corresponding to requests. Consequently, if one requested the inscription of “Salut” on an ancient pedestal, the resulting writing will differ from the required text. However, this problem is now resolved thanks to DALL-E 3.

Generation of scientific diagrams

The new generation of texts for images could inspire you. For example, it would be well suited to creating scientific diagrams. However, OpenAI points out that artificial intelligence can give inaccurate results. However, if you provide all the necessary information, schema creation should, in principle, work accurately. Additionally, you don’t have to provide all the information, as ChatGPT will automatically complete your message before sending it to DALL-E 3.

In reality, drawing a diagram will not work entirely well. The example above is quite specific, although normally simplistic. Two rectangles should have been drawn, connected by a link, and in each table, member data should be entered. There, DALL-E 3 clearly did not do that, and the result is not very convincing.

Artificial documents: this is a categorical refusal!

An important point that the system is capable of doing is the generation of artificial documents. This still follows from textual generation, but the difference here is that the generator is very good for this document generation. Unlike scientific images which were inaccurate, the generated documents can fool people. The model used is therefore capable of doing this. However, for ethical and moral reasons, OpenAI has decided to prohibit this possibility.

Inspiration of artistic styles, protection of intellectual property

DALL-E 3 likely uses various artist creations in its source data to be properly trained. Because of this, it can reproduce images in the style of a specific artist. However, this can cause intellectual property issues, both regarding generated images and those used for training. However, in order to avoid possible legal action, DALL-E 3 restricts the creation of images based on the style of artists, particularly those still living, with a handful of exceptions. These include world-renowned artists such as Picasso.

The new capabilities of DALL-E 3

Photorealistic images, the spread of fake news is coming?

The images produced by DALL-E 3 are becoming more and more realistic, making it difficult to distinguish from real photos. It is difficult, if not impossible, to recognize an image created by artificial intelligence with just a glance. If we take up a subject that shook France a few months ago, we can have fun creating beautiful fake photos. So let’s take the subject of pension reform, simply asking for a generic photo illustration. The smartest (or perverse) minds will think about the possibilities of creating false images, of a demonstrator breaking works of art, of police blunders or many other false information.

Caricatures and cartoons

One of the graphic styles of the DALL-E 3 generation that has come out the most on Twitter is the cartoon. This style is very well generated and reproduced by artificial intelligence, with very impressive results. Sometimes the texts are not entirely operational, but in the majority of cases the results are very good. So, with a general theme, you can ask for examples of caricatures around a subject. DALL-E 3 (via ChatGPT) will launch the generation of different images from different prompts, which will diversify the responses.

DALL-E 3 is therefore a very interesting tool, which will become part of the habits of many users. Whether to illustrate a PowerPoint presentation, to create a meme for a friend, or for articles, the image generator is very strong. The most interesting thing now is to know what OpenAI is preparing to create a DALL-E 4 that will be even better. To be continued…