Dall-E, it can generate many imaginative images from the description, even some peculiar images, how did they actually create this kind of dataset to train this AI , because there is not much of that kind of data which include weird images and descriptive text, how did they create this massive dataset. Does anyone have any idea?

If you have no idea what I am talking about, please refer to this link: https://openai.com/blog/dall-e/.

  • $\begingroup$ Have you read the article that you're linking us to, i.e. openai.com/blog/dall-e? It says they are using a smaller version of GPT-3. They even say that they use a dataset of descriptions and image pairs. I didn't fully read it, but this already looks like a partial answer to your question. So, what's your question? $\endgroup$ – nbro Jan 21 at 12:09
  • $\begingroup$ yes, it gets clearer for me when I read deeper into it, and I don't quite understand how this data is obtained and reformatted for training to make this AI working. thank you $\endgroup$ – Gavin Jan 21 at 14:30
  • $\begingroup$ Ok, then edit your post to ask specifically that (although that will actually and partially invalidate the current answer below), i.e. how exactly is the dataset of text-images created? How many images does it contain? What are the sizes of these images, etc? Or whatever you are really insterested in. Because, otherwise, this post doesn't make much sense, given that the answer to the current question can be found in the article that you're linking to and suggesting us to read in order to be familiar with Dall-E (which you supposedly were). $\endgroup$ – nbro Jan 21 at 14:32
  • $\begingroup$ sorry, thank you for advising $\endgroup$ – Gavin Jan 22 at 0:29
  • $\begingroup$ That edit didn't really improve the post. Please, read my comments above again. $\endgroup$ – nbro Jan 22 at 12:22

DALL·E is a 12-billion parameter version of GPT-3 trained to generate images from text descriptions

should be the same data they used to train the GPT-3

  • $\begingroup$ but what did they use as output for their input text $\endgroup$ – Gavin Jan 21 at 5:26
  • 1
    $\begingroup$ they are mapping text to images. so their dataset should have a pairs of text and image and their model trained on mapping these pairs. $\endgroup$ – Thulfiqar Jan 21 at 10:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.