I recently came across an interesting interview in Venture Beat with Aditya Ramesh, the inventor of Open AI's DALL-E. In it, Ramesh claims to have been surprised by the remarkable popularity of his extraordinary imaging app, and I don't think anyone can blame him. A mere two years ago the very concept of an AI imaging app was unheard of, and there was no way of knowing what a sensation DALL-E would become when once released. Still, the promise it held out was irresistible - that one had only to type in a text description in order to see an pictorial representation generated almost before one's eyes. No wonder so many, including myself, rushed to put their email addresses on the waiting list in order to be among the first to use it. And no wonder Open AI is now considering a tender offer that would place its value at $29 billion.
Of most interest to me was Ramesh's description of how images are generated. He noted that images are generated from scratch, beginning with a blurry outline, rather than by cut & pasting from the huge inventory of scraped images in its database. This was of particular importance to me since I strongly believe in respecting other artists' intellectual property rights just as I would want my own to be. I feel I can rest easier now knowing that I am not inadvertently appropriating anyone else's images or parts thereof.
As far as DALL-E 2 itself is concerned, I've worked with it and admire what it can do, but I've actually found the images generated by Stability AI's Stable Diffusion to be more aesthetically pleasing. I realize, of course, that this is a very subjective choice but I nevertheless find myself working far more often with Stable Diffusion than with DALL-E 2.
No comments:
Post a Comment