Why are the images created by Dal-e AI so ugly (and ridiculous)?

If you chat a bit on the internet, you wouldn’t miss it: Dal-e’s creations are everywhere. Distorted faces, very intense colors, even distortions that cause discomfort … sometimes horrible vision that results from artificial intelligence (AI for short), a supposedly intelligent computer calculation program, or in any case that drains water from text to image. Canvas. And in this case, there is no limit to your imagination: the more precise your formulation, the more AI will create an image with the requested elements. And often, it’s too ugly and too funny. To try, it happens here, and you only have to write one sentence, especially in English, and let the machine run.

Elmo – Dal-e Mini screenshot as a guest star in the Friends series

The Dal-e-Mini program is the work of Boris Daima, and co-founded by Elaine Musk, an OpenAI company specializing in artificial intelligence. Dal-e Mini is a program responsible for creating a textural proposition image, in the 12 billion parameter version of the GPT-3 language model, also created by OpenAI. This algorithm is based on the so-called natural language, that is, the way we communicate as human beings as opposed to formal languages ​​like computer languages. Unveiled in early 2021, the AI ​​program has been a wild success on Twitter in just a few weeks. A lot Twitter, Discord or Reddit thread The images produced by Dal-e-Mini are dedicated to Originally, one of OpenAI’s main goals was to “use language models to better understand everyday concepts that people use to understand things.” MIT Technology Review.

Why is the generated image so ugly?

Much like artificial intelligence, Dal-e learns through it Acquiring deep knowledge (Or deep learning in VF): Through the processing of so-called disorganized data (words, images, human language, etc.), the program goes through a so-called “neural network” algorithm that mixes different levels of different calculations. Apparently, a computer program tries to create a link between a sentence and its meaning in the case of dal-e in a few pieces of data. For example, we asked Dall-E to integrate Emanuel Macron on the show. Dance with the stars. The results are amazing.

Emanuel Macron to dance with the stars
Emanuel Macron in Dancing with the Star – Dal-i Mini Screenshot

But as we can see, the results are often more interesting than real. Although Dal-e has become one of the memes, above all it makes it possible to keep in view the supposed superiority of artificial intelligence over the human brain. If AIs develops faster, more or less, we are far away I’m a robot. For now, the Dal-i Mini still serves as a threat to the (limited) power of AI and their ability to shape our future … but without relying on its evolution, Dal-i2.

Dalida fashion cover metal
Dalida Way Cover Metal – Screenshot Dalida-e Mini

Specter of DAL-E2, or more precise AI than before

Last April, more than a year after the unveiling of the Dal-e Mini, OpenAI announced the presence of the Dal-e2, claiming that it could create a more realistic image than ever before from the text description. According to its creators, the Dal-E2 has been described as a model that “can create original and realistic images and works of art from textual descriptions. It can combine ideas, features and styles.” Since then, There are beta tests with handpicked users who publish some ultra-realistic results of AI application in a single thread.

Because unlike its “mini” version, which offers irrational combinations, Dal-E2 now understands the relationship between objects, but also retrieves and edits images in a realistic way from a written description. According to its creators, AI will be able to replace even part of the image with a mixture of other automatically generated images. Cold and Surprise: Dal-i became an image maker, and was able to go further (and above all faster) than the average person.

Due to this fact, OpenAI has strict rules. The party has removed all violent content, installed filters and a policy against any kind of nudity, conspiracy or political content. This is because the developers of OpenAI mentioned on their site, “Without adequate protection, models like Dal-E2 can be used to create a wide range of misleading content”. In times of dipfeck and fake news, OpenAI prefers not to take any risks … and probably keeps in touch about its mini version and its ugly pictures.

Leave a Comment