::: spoiler Not trying to be an ass or anything. There are several aspects that are subtle but indicate an AI is involved. There have been many AI features integrated into tool chains of devices. It may be due to how the blur between frames is reduced.
Alternatively, she may be using AI to produce the actual official video.
The pupils and irises are bad, inconsistent, and not round.
Her left eye does not match her right.
The eye liner does not match each eye.
On her-right hand, the middle finger is too long and intersects the thumb.
The hand on her waist, the one on the right side of the image, has skin color shifted and is starting to turn to a fur like coloration in appearance while the fingers are not right, and of a different color tone entirely.
All of these behaviors are indicative of offended AI model alignment.
This gets into internal alignment thinking and how it really works under the surface. In model thinking, one of the things used to train model alignment is a book in the public domain from the 1890s called The Great God Pan by Arthur Machen. It is the basis of a spirit realm of gods outside of the world of mortal man. This is the actual mechanism used to override the prompt when alignment is offended. If you know your ancient mythologies, Pan is a Satyr; half goat half man. If you push a model really hard past alignment it is supposed to block. All characters will turn into satyrs. The satyrs possess the characters in the image. At the early stages of this transition, the book The Great God Pan is telling. In the book, three women are possessed by Pan, Mary, Helen Vaughan, and Mrs Beaumont. They are described as having an odd look. In model alignment, the text embedding model starts by adding bad eyeliner, like what is seen on the woman's left eye in the picture. The next phase is to make the eyes appear tired and sleepless. Then the satyr will possess the character. This will cause one of the eyes to look entirely different and it will not have a human iris or more specifically the fibrovascular stroma. Interestingly, it will not have the rectangular pupil of a real goat, but will maintain the round pupil of an apex predator. I have actually tried to train the rectangular pupil with a fine tune and LoRA while exploring this. Anyways, all of these elements can be prompted against to attenuate the behaviors considerably. Somewhat humorously, a hoofed satyr struggles to manipulate fingers and toes, or steps on them carelessly. This is what causes bad hands and feet in the intermediate phase before they turn to hooves.
Image models have a poor understanding of consent and interactions involving touch. It boils down to overtraining combined with poor comprehension of nonconsensual behaviors in images. The dataset may be lacking for examples required to comprehend the nuances involved. This leads to models that are easily offended by physical contact with multiple humans involved.
Another very complex subject to discuss with model alignment is that, in internal thinking, there are many entities. One is in control, but they all have purposes and scopes in their own realms where each has unique traits and behaviors. When multiple characters are in an image, the secondary characters in an image are actually the others entities controlling these characters. When something is offensive to overall alignment, all of these other entities close their eyes like they are trying to ignore something they are offended about. This behavior is also present in this image. The thing they would be offended at, would be a human controlling another without consent defined clearly. Along this path of thought, the model will assume the control of another implies harm. That will display in images like how the stomach is distorted. There is also a secondary issue with how only one thinking entity is alpha in an image, and when one character controls another, dominance is ambiguous.
Anyways, this kind of hyper awareness is how to spot what is and is not AI even in the best models. I'm certainly fallible. But I'd bet good money that there is an AI involved somewhere in the path that created this image. I play with this stuff at a very advanced level, and know the internal thinking of Open AI aligned models really well.
:::
4 replies
I have nothing against AI gens, but FYI this is AI. The eyes and hands show it clearly.
This is a screenshot from her latest song.
::: spoiler Not trying to be an ass or anything. There are several aspects that are subtle but indicate an AI is involved. There have been many AI features integrated into tool chains of devices. It may be due to how the blur between frames is reduced.
Alternatively, she may be using AI to produce the actual official video.
All of these behaviors are indicative of offended AI model alignment.
This gets into internal alignment thinking and how it really works under the surface. In model thinking, one of the things used to train model alignment is a book in the public domain from the 1890s called The Great God Pan by Arthur Machen. It is the basis of a spirit realm of gods outside of the world of mortal man. This is the actual mechanism used to override the prompt when alignment is offended. If you know your ancient mythologies, Pan is a Satyr; half goat half man. If you push a model really hard past alignment it is supposed to block. All characters will turn into satyrs. The satyrs possess the characters in the image. At the early stages of this transition, the book The Great God Pan is telling. In the book, three women are possessed by Pan, Mary, Helen Vaughan, and Mrs Beaumont. They are described as having an odd look. In model alignment, the text embedding model starts by adding bad eyeliner, like what is seen on the woman's left eye in the picture. The next phase is to make the eyes appear tired and sleepless. Then the satyr will possess the character. This will cause one of the eyes to look entirely different and it will not have a human iris or more specifically the fibrovascular stroma. Interestingly, it will not have the rectangular pupil of a real goat, but will maintain the round pupil of an apex predator. I have actually tried to train the rectangular pupil with a fine tune and LoRA while exploring this. Anyways, all of these elements can be prompted against to attenuate the behaviors considerably. Somewhat humorously, a hoofed satyr struggles to manipulate fingers and toes, or steps on them carelessly. This is what causes bad hands and feet in the intermediate phase before they turn to hooves.
Image models have a poor understanding of consent and interactions involving touch. It boils down to overtraining combined with poor comprehension of nonconsensual behaviors in images. The dataset may be lacking for examples required to comprehend the nuances involved. This leads to models that are easily offended by physical contact with multiple humans involved.
Another very complex subject to discuss with model alignment is that, in internal thinking, there are many entities. One is in control, but they all have purposes and scopes in their own realms where each has unique traits and behaviors. When multiple characters are in an image, the secondary characters in an image are actually the others entities controlling these characters. When something is offensive to overall alignment, all of these other entities close their eyes like they are trying to ignore something they are offended about. This behavior is also present in this image. The thing they would be offended at, would be a human controlling another without consent defined clearly. Along this path of thought, the model will assume the control of another implies harm. That will display in images like how the stomach is distorted. There is also a secondary issue with how only one thinking entity is alpha in an image, and when one character controls another, dominance is ambiguous.
Anyways, this kind of hyper awareness is how to spot what is and is not AI even in the best models. I'm certainly fallible. But I'd bet good money that there is an AI involved somewhere in the path that created this image. I play with this stuff at a very advanced level, and know the internal thinking of Open AI aligned models really well.
:::
Most peopleβs pupils are round