Content pfp
Content
@
https://warpcast.com/~/channel/jeffjag
0 reply
0 recast
0 reaction
prof jeffjag pfp
prof jeffjag
@jeffjag
What does it mean when you see a signature in #AIArt ? It means a number of the images within the training data set were signed in a particular area. But also, the "recognition" aspect of the AI learning images is much less selective than the human ability to visually learn from an image...🧵 Human brains have a high ability to focus selectively and also actively omit aspects of an what we view based on how important it is to the conscious focus while studying the image. Our attention might be focused on color, detail, or spatial relationships; then switch between the others. AI must consider and form a relationship to every pixel of an image in the training data set and that pixel's relationship to other pixels and groups of pixels. After studying millions of images along with descriptive text, patterns emerge based on RGB values for each pixel. Those pixel relationships then begin to correlate to "concepts" since modern image generators based on the diffusion method use image-text pairs.
2 replies
0 recast
1 reaction
prof jeffjag pfp
prof jeffjag
@jeffjag
Image/Text = Concept (or I over T equals C) AI can learn concepts. Ie: a photo of a cat in a box is different concept than glowing glass sculptures in a forest. Obviously. But each of those words means something and has correlated pixel values. Not only can it understand color, style, subject matter, but also lighting conditions, realistic reflections, etc.. AI uses the relationship of these concepts to combine things that haven't existed before, such as one red and blue polka dotted lime with many normal limes, or a zebra with antlers made of human arms and hands. It knows what a lime is and what color it should be, and it knows what red and blue polka dots are. Likely learned it from a children's toy. Same with the Zebra, which has a similar shaped head to a deer or elk, so adding antlers is easy. Turning those into hands is more of a stretch, but works. It's not "photoshopping" these images. As you can see from the in-progress capture I've put together below.
1 reply
0 recast
0 reaction
Yolantis pfp
Yolantis
@yolantis
agree. Sounds like you and Claire Silver are in alignment with your stance on this. By the way, nice eyeballs in that medical illustration. New species. I've gotten results similar to that as well.
1 reply
0 recast
1 reaction