>>72481438I got some new training tricks from the other place that got overlooked.
>The neural network associates words to image components. If you give the AI an image of a single finger and tell it it's the ring finger, it can't know how to differentiate it with the other fingers of the hand. You might give it millions of hand images, it will never form a strong neural network where every finger is associated with a unique word. It might eventually through brute force, but it's very inefficient.>Here, the strategy is to instruct the AI which finger is which through a color association. Two identical images are set side-by-side. On one side of the image, the concept to be taught is colored.>In the caption, we describe the picture by saying that this is two identical images set side-by-side with color-associated regions. Then we declare the association of the concept to the colored region.>Here's an example for the image of the hand:>"Color-associated regions in two identical images of a human hand. The cyan region is the backside of the thumb. The magenta region is the backside of the index finger. The blue region is the backside of the middle finger. The yellow region is the backside of the ring finger. The deep green region is the backside of the pinky.">The model then has an understanding of the concepts and can then be prompted to generate the hand with its individual fingers without the two identical images and colored regions.reddit aolvxz
