Shadow^Mist posted at 6:22 PM Tue, 23 January 2024 - #4480719

Qaz, I don't mind at all. Please go ahead and post on-going improvements. As a side note, I found that getting a single figure to look like the original is much easier than getting a scene of several figures with a background to look like the figures. I'd appreciate any insight you discover re that. 

OK so from working with this for a while, it is clear that the AI works great to an extent but it has limitations. Its not good with hands. I'm told that you can overcome this with in depth negative tags, but not found that yet. If there are small intricate things in the picture such as faces the AI struggles. I think since this part of the program is trying to enhance what is already there it struggles when it has low level of pixels to work with. Additionally the data set it has is going to consist mainly of people looking forward towards the camera, and struggles with unusual angles. A scene where stuff is happening is going to be hit and miss. 

So this is a render I did 10 years ago in firefly and I was happy with it at the time, but the Superfly versions would have been better. I did this for Halloween.

And this is what I eventually came up with:

Now it was originally messed up. So basically I cheated. I thought what if you cut up the individuals into separate pictures and enhanced them. Much better. As you have seen large pictures of faces come out pretty well. I had problems initially with the Megan Fox like vampire as she was on her side, so I thought what if I turned her 90 degrees so she is upright and pressed enhance. Yep that worked. Cut and paste them into the same picture. Make a few other corrections such as two of the hands and voila ! And to think this AI is just getting better and better.