Recent searches
Search options
Administered by:
@penguin42 Perhaps this trick is common enough that the model was able to learn it from the training set 
The larger image shows that Dr. Arroway is actually wearing a thin catsuit.
@penguin42 It seems DALL-E has most trouble with unusual placement of objects, such as "person behind server rack".
@penguin42 Actually, it does a good job if you ask it to do only one thing at a time:
@penguin42 This is the prompt created by #ChatGPT for #DALLE:
"Illustration: In a neon-lit futuristic datacenter, a woman resembling Dr. Ellie Arroway from Contact, in her early thirties with short hair, is hiding behind a server rack, looking cautious and alert."
"Illustration: In a neon-lit futuristic datacenter, a woman resembling Dr. Ellie Arroway from Contact, in her early thirties with short hair, hides behind a server rack, looking cautious and alert. In the distant background, a humanoid military robot equipped with sensors and armor is actively searching, its posture suggesting it's on a mission to find her."
As soon as we add one more element, things start falling apart:
"Illustration: In a datacenter, a woman hides behind a rack. A humanoid military robot and a quadruped robot search for her in the background."
Result: Dr. Arroway takes her puppy to work 
@codewiz @penguin42 I wonder if they're using the image analysis to train ChatGPT to generate better prompts.
@jeeves @penguin42 That's what I would want: loop the two models until they figure out how to generate a good illustration for my new sci-fi novel.
ChatGPT knows *exactly* what the story is about - being the ghost writer for the entire thing - and surely could come up with a decent idea for its cover as well 
It's sad that I have to micromanage these two bots to make them work together 
