Create realistic images from text and in a few seconds, that’s what it proposes DALL-E. OpenAI’s artificial intelligence model has proven to be truly spectacular; Y DALL-E 2 It is on everyone’s lips because it is capable of generating realistic images and art from a description in natural language, successfully understanding —and with originality— even the most complex and twisted orders that we can imagine. It is, without a doubt, a technological marvel; but it is not available to everyone. Until now, only a very limited number of people have been allowed access, but fortunately there are projects that seek to remedy this. Among many stands out DALL-Emini.
DALL-E mini is a reproduction of DALL-E, but open source. What this tool proposes, developed by Boris Dayma, is that anyone can create their own original images from a text description. And being open source, any developer can access to the repository on GitHub and start training your own AI model.
According to its creator, DALL-E mini’s training has consisted of exposing the AI to millions of images available on the web with their respective annotations. In this way, as it “absorbs” the concepts of each of them, it learns to generate new images according to what is requested through a text order. It is worth clarifying, however, that the project still in training stagealthough the first results are interesting.
Some of the concepts are memorized as they may have been seen in similar images. However, you can also learn to create unique images that do not exist, such as “the Eiffel tower landing on the Moon”, by combining various concepts.
Boris Dayma, creator of DALL-E mini
DALL-E mini offers good results, but still has a way to go
Let’s be clear, the DALL-E mini has spectacular potential and the results it offers, given the descriptions we present to you, are good. But don’t expect it to work at the same level as OpenAI’s artificial intelligence, at least not yet.. And it is logical that this should happen, since we are talking about initiatives conceived with very different scales and resources.
That gives the DALL-E mini even more value, despite the problems we may run into when testing it. Below you can see the results we got with “Homer Simpson playing football” (Homer Simpson playing football). If we say that the results they are surrealwe will probably fall short.
Another interesting point is that text commands do not necessarily have to be in English. We obtained very similar results in Spanish, in this case with “Homer Simpson playing tennis“.
And here you can see what was generated with “a dog using a computer” (a dog using a computer).
While the descriptions get even more complicated, the DALL-E mini still doesn’t hit the nail on the head with imaging. When we entera horse riding a motorcycle under the rain” (a horse riding a motorcycle in the rain), it has become clear that he still needs to polish a few details of the final result.
We come back to the same thing: the potential is there. DALL-E mini is able to recognize the pieces that are part of the “puzzle” that we propose in each text description; once artificial intelligence tightens the screws to unite everything in a quality image, it will take the definitive evolutionary leap.
The real power of DALL-E mini is that it is an independent and open source project that is sure to get better as more people use it. If you want to try it, you can do it at this link. Please note that you may receive frequent traffic alerts, and the final images may not yet be saved in high quality.