... plus it's absolutely addictive. Here's an image it created of Bobbylee Johnson and his old pickup truck, the Rebel Rambler. He's standing in the parking lot of the McDonald's in East Weevil, Indiana.
ChatGPT is now bound to the AI image-generation program, Dall-E. The interface, being ChatGPT, is light years ahead of the horrid chat-session interface for Midjourney. You can go back and forth with ChatGPT creating a prompt for it to create an image before actually doing so. The AI is helping you use the AI.
Having said that, the AI gets confused as soon as there are more than 2 characters specified in a scene. It can't figure out the positioning at all and the costumes get blended sometimes as well. I'm still playing around with Arthurian fanfic and the prose and dialog are much, much better than they used to be. Sometimes, if there's a part of the story for which I want an illustration, I ask ChatGPT to generate a Dall-E prompt to create it. If there are one or two characters, Dall-E typically nails it in a few tries. More than two characters and you might as well just quit.
Sometimes, it gets close, but refining it to get exactly what you want is beyond frustrating.
In that scene, the knights were wielding daggers and their armor should have shown signs of a furious duel. The blue knight was supposed to be dead. That image was as close as I could come to it.
When Dall-E generates an image, it tells you the modifications to your prompt that were used to create it. When you ask it to generate, it will typically create four, performing slight alterations to your prompt. You can then pick your favorite and ask for modifications to it. For that scene, Dall-E was convinced that the blue knight was dead and they had been using daggers at the end of the fight. It also thought their armor and surcoats were torn and soiled by combat.
Further refinements of that image resulted in bits of armor all over the meadow, extraneous characters in the background, the princess wielding a sword or wearing armor and all other manner of excrescences just short of space aliens and circus elephants.
So far, two characters in a scene can get you the image you want in a few iterations, but beyond that, Dall-E cannot figure out the geometry, costumes or expressions to save its life. Either that, or I don't know how to write the prompts properly. That's entirely possible.
Having made all those complaints, the thing is an absolute marvel. I love seeing my characters come to life in illustrations. For Bobbylee, I wanted something much closer to Bo Duke, but I ran out of time last night to continue refining it.
 |
| John Schneider as Bo Duke. |
Dall-E also has some other interesting limitations. It refuses to create images based on historical figures or celebrities. I guess this is to prevent deep fakes. I asked it to generate an image of Robert E. Lee wearing sunglasses and smiling and it wouldn't do it. It also is sensitive to what it displays. I'm not allowed to ask for bloodstains in my Arthurian images nor a rebel flag for the Rebel Rambler.
Oh well. What do you expect for $20 a month?
That is a pretty realistic image of a beaten blue blooded lizard knight...
ReplyDeleteIt sounds like you are now hitting the point where the chatbots are roughly equivalent to hiring an artist or a ghost writer to draw/write for you, which is indeed a pretty big deal.
ReplyDeleteOf course, the stumbling block here is that ghost writers and cheap artists are notorious for producing uninspired, unmemorable dreck. I recently saw a quote from Shel Silverstein to the effect that nobody knows what it is in a particular piece of art that "makes the magic happen", and until someone does figure that out, we aren't going to get anything really special out of the AIs except by accident.