Possibly combining ChatGPT and DALL-E to make a powerful art teaching program (maybe)

It’s not really an “inverse”, they are standalone products using different technology.

GPT-4 was trained to be multi-modal with computer vision, so it has an understanding of constructing images in the same way that it constructs language, by their likelihood of appearing in a certain way based on inputs, with machine learning refining the internal weights. Parts of imagery training may be relying on tagged data, so when you have 1000 “dog noses” it can synthesize how other images of noses would activate the same neuron and semantic meaning. A very advanced version of OCR text recognition.

Dall-e is more input = output, and is more uncertain. It will make abstraction and uncanny valley and impossible visuals as well as the occasional impressive piece from text. It is not the leading edge of image generation, others like Midjourney have comparable qualities.

Let’s test the “inverse”. The AI initial description of the student art. Then I pick maybe the worst of the four (although another looks like a corpse)