Lately, I've gotten a lot of horrible deformities, not just the old 'six fingers' on a hand.
Here are a few from today. I saw much worse and didn't save them. One was a woman with two heads and a third arm separating them, but then a third head where her vagina should have been. My prompts were fairly simple and normal, so I have no idea how the AI got things so horribly wrong.
Anyhow, here are a few from today, from perchance.
I waited a few days to comment here, as I don't intend to bump the post.
I was thinking that this might open a discussion about this problem, and to know how severe this problem may be, and even perhaps sharing some of your own franken-AI images.
With technology that can accurately illustrate a messy substance flowing through hair, it seems so strange to me that it can't get 'five fingers on a hand' correct, or worse, the basic form of a human body. But we know AI is always evolving and this will one day be a thing of the past.
In the meantime, any comments?
11/18/23, 10:52am: This post won't bump the thread to the top.
these images will give me nightmares.... most remind me of the "the thing" mid transformation like you caught it before it could get the human features right... another one looks like coneheads
ive never had it this bad from bing but it does go weird sometimes.
They do look like melting cakes, and the conehead cracked me up.
Hopefully they will straighten things out more so we will eventually have the right number of fingers, toes, arms, heads, etc. and in the right places. But I'd say a good 80 to 90 percent of images look fairly natural.
11/18/23, 10:52am: This post won't bump the thread to the top.
I think stable diffusion struggles with landscape images and lying poses. I struggle to get good results. There are so many orientations and angles. Portrait photos are quite easy. Sometimes I experiment with 768x512 (3:2) and get bad results like this. I get best results with 512x768 (2:3). I get mostly good results with 768x768 or most (1:1) images. I usually upscale by 2.5x.
I tried to use ControlNet with an OpenPose editor. I am able to get specific portrait results. I tried landscape with lying poses and I still get bad results.
I don't have any suggestions to get better results for horizontal poses. It is very difficult to describe a pose. I usually stick with standing, sitting, and kneeling.
DALL-E3 is able to get much better results with lying poses right now, and the hands are much better. It seems to struggle with faces and eyes sometimes.
Sometimes I get a really funny result. I was trying to create an image of a woman in a doggy style pose with green goop on her butt, and a dog leash around her neck. It created a woman standing next to a dog, and the dog is covered in green goop. I don't know if I should share it. The dog does not have a tail. It is clearly a dog, sitting, no tail, and covered in green goop. I am afraid to show a depiction of a messy animal due to reactions it might get from people thinking it is showing something cruel.
Sometimes I get a really funny result. I was trying to create an image of a woman in a doggy style pose with green goop on her butt, and a dog leash around her neck. It created a woman standing next to a dog, and the dog is covered in green goop. I don't know if I should share it. The dog does not have a tail. It is clearly a dog, sitting, no tail, and covered in green goop. I am afraid to show a depiction of a messy animal due to reactions it might get from people thinking it is showing something cruel.
I had a similar thing happen when I was running through yoga poses and tried downward dog. Bing still has a problem with context occasionally like this but some of the issues are due to needing to work around the filter so much. It would be much easier if you didn't have to ask for something like... "Sitting on knees, hands forehead on floor with rear pointing up"
Her foot contains one big toe and about 25 little toes. Singing 'this little piggy' would take all day. I had a horse-head pop up on another one, similar to the cow head you had, which ruined an otherwise great image.
Love you, too
11/18/23, 10:52am: This post won't bump the thread to the top.
Very exciting experiments! Now it looks like horror movies from horror films - in good quality
I think that if AI learns to do all these sketches well, then: 1. demand for WAM models may disappear 2. This can create danger around all people, because every picture could be created with the participation of another person without his knowledge (like compromising or blackmail measures) I think this can all go far if it starts working perfectly
This appears to be for entertainment purposes only at this time.