I messed around with NJ a little more this morning. While it’s certainly better than straight MJ, it still fails a considerably high percentage of the time when generating complex scenes. By fail, I mean that anatomical elements are borked, background characters are completely out of scale, etc.
Let’s take a look at some of my results:
These ballgown/ballroom scenes are all fairly simple and the AI manages without issue. However, prompt for a more complicated scene and you see how quickly things can fall apart:
I can’t tell you how many (dozens) re-rolls I attempted to get one or two good results. Now, Stable can’t necessarily do better, so I’m not really shitting on NJ/MJ here. Perhaps my expectations grew unrealistic after my rather successful- but admittedly simple- initial results.
Still, I had time and the willingness to persevere in order to get some plugsuit-ery:
I felt that these two could tell a bit of a story… As our companion changes, so do we:
Next up are a few casual scenes. The single subject and simple background is where the AI really shines and it only took a couple re-rolls to get results that I genuinely liked:
I wasn’t intending to share a million images or write much more than what I have, so I’ll leave things off here… With Asuka’s Parisian vacation!
