DALL-E, the new OpenAI model

How innovative is it? It uses two things already well known in the world of AI, natural language processing, with a version of its GPT-3 model, and the generation of realistic images. This gives impressive results, as can already be seen on its website (https://openai.com/blog/dall-e/), that with only a text describing the model can generate an image. It's the first time that an AI has achieved something like this, the models that already existed like StackGAN were limited to one type of image, in that case, birds.
A few months ago they had already managed to get the GPT-3 model to generate images based on a part of them, processing it as text, that is, as a unidirectional sequence of characters, unlike the convolutional neural networks that are usually used for this type of problem, which work with two-dimensional matrices (https://openai.com/blog/image-gpt/). DALL-E uses the same idea but scaled up in a big way since it has 12 billion parameters.

Posted Using LeoFinance ^Beta