Some time ago we learned that DALL-E 3, the next generation of image-generating artificial intelligence from OpenAI, he was already in the test. Although we expected its debut to happen before the end of 2023, we have had no evidence of this until now. And since generative models are on hold due to regulatory moves happening around the world, we can’t take it for granted either.

However, it seems that the creators of ChatGPT have decided to go further, and so just a few minutes ago message in your X account, OpenAI announced the release of DALL-E 3a new version which, based on what they’ve shown so far, seems to be a significant improvement over DALL-E 2, the version currently available and which, as you may remember, we already tested about a year ago. which left a somewhat bittersweet taste in our mouths.

Recall that DALL-E 2 technology is currently available for free to Microsoft Bing users, as well as to Microsoft Designer users. And something interesting when you see the way DALL-E 3 allows you to work from an original idea with step-by-step challenges is that OpenAI seems to have taken good notice of the way image generation was implemented in Microsoft’s chatbot.

OpenAI puts a lot of emphasis on improving DALL-E 3 compared to previous versions (which can also be attributed to other image generation models). If you have used a service of this type, you have surely seen it The model ignored some of your quick instructions, which happens more often when we work with more complex requirements. This point is so important that when accessing the website about this new model, this is the first thing we can read on it:

«DALL·E 3 understands much more nuance and detail than our previous systems, allowing you to easily translate your ideas into exceptionally accurate images.«.

And if that wasn’t enough, they add:

«Modern text-to-image conversion systems tend to ignore words or descriptions, forcing users to learn fast engineering. DALL·E 3 represents a leap forward in our ability to generate images that exactly match the text you provide.»

If they have indeed achieved this, either with the first prompt or subsequent indications regarding the original image, we are undoubtedly talking about really significant progress and this will significantly improve the working experience with this type of tools.

In fact, in this regard, DALL-E 3 seems to be OpenAI’s first response to what I believe to be the fleeting figure of rapid engineering (which I use to emphasize that the use of the term “engineering” seems to me redundant in every rule). It is clear that the companies responsible for this type of service aim to make it as easy as possible for users to get the desired results. Therefore, This new generation of DALL-E could mean the beginning of the end of this new “profession”.

The image you can see below the bottom paragraph shows you another of the big news of the DALL-E 3 compared to its predecessor, namely is now able to generate readable text on imagesand not only when reproducing the texts we give in the prompt, but also when generating our own text from what we give in the item, from which we can deduce that it uses a modified version of GPT-4 that will be integrated into the DALL-E workflow.

According to the first links in this regard text generation, at least for now, is not perfect, which is the same thing we could see a few weeks ago in Ideogram. However, this is less of a problem in this case because, as I mentioned earlier, the result of the first call can be tuned by iteration. So DALL-E 3 aims to be the model that will finally provide a tool for creating images with text that offers the required quality in both types of content.

Whenever we talk about generative models, it’s important to consider everything that goes with it security, respect for intellectual property and, particularly in vogue recently, the generation of false images about real people. OpenAI has taken these points into account, as we can see on the new model website, with a set of measures aimed at mitigating risks and offering solutions.

Firstly, and in line with measures already taken in previous versions, OpenAI limits DALL·E 3’s ability to generate violent, adult and hateful contentto which he adds a mitigation for reject requests that request images of public figures as the name suggests, in addition to improving its ability to detect content that can be used to generate disinformation, manipulate public opinion, etc.

In addition, this new generation of the model was designed inreject requests that request a painting in the style of a living artistin clear response to the numerous criticisms generative models have received for mimicking styles, voices, etc. Additionally, a space has also been added for creators to tell OpenAI that they want their images excluded from training future image generation models.

As for her debut, here’s what we can read on the service’s website:

«DALL·E 3 will be available to ChatGPT Plus and Enterprise customers in early October. As with DALL·E 2, the images you create with DALL·E 3 are yours and you do not need our permission to reprint, sell or trade them.»

Interestingly, access to DALL-E 3 via API is not mentioned, Although it is true that there is no reference in its marketing to the specific version of the model it provides access to. However, we’ll have to wait a little longer to see if the initial access extends to those using the API to use OpenAI services.

And now the question I’ve been wondering since the announcement of the DALL-E 3 launch, Will it be somehow related or will he have any participation in the Microsoft event tomorrow Thursday? Recall that when GPT-4 was announced, Microsoft immediately confirmed that the new model is already present on Bing. Granted, yes, image generation is more limited in scope than text generation and chatbots, but it could still be a service with huge scale, and given Microsoft’s huge involvement in OpenAI, it has quite a lot of potential. It makes sense to think that they want to add its technology to their products and services.