User blog:Dai4lyf/Creating an AI to generate friends for myself

SEE PREVIOUS BLOG POST (no link)

Im going to make an AI that will create 15 faces that will be my friends because im lonely yay

The training data set will consist of all the mii faces taken from this wiki and all the admins faces (excluding pictures of Admin's profiles that aren't set (like User:616Kate616), pictures of the given profiles (fish, cookie, earth, dog, flower, or grey portrait on whitish background) (like User:Végégon) and pictures of admins with no miis or more than one mii (like User:Unknown Channel or User:YoungArtist79 or User:Explodey Boi))

And afterward the AI will have to decide where the faces in all the portraits are and the background and the shirts. To find the faces the AI has to use a filter. But the reason why I excluded certain admin's pictures is because if Unknown Channel is loaded into the Data set, the AI won't know which face to detect, so it detects both and the faces the AI generates will look static (because of the number of faces on screen at once) And also messy faces (like User:Zerousage) because the AI will have a hard time seeing the mouth as it is covered up by the moustache. And also non-faces like User:Jonasabbou13 because the AI can't detect a face in the photo. So that's that.

And after that the AI will have to guess the average of all faces displayed on the Data. BUT WAIT! some miis have a unique hairstyle, or unique features like Mole placement, Mole size, Eye shape, Eye size, Eye Rotation, Eye parting, Lip Shape, Lipstick Color, Facial Hair, and Mouth Placement. For Example, Cole has a unique eye shape, and Gwen has a unique nose shape. Ian has the least eye distance and Pablo has the most eye distance. Mónica has a unique hair style, and Greg has a unique hair style. Along with the obstacles with the Admins Faces, Unique parts tell the AI that hardly any people like this in real life. Now, we will accept Admins with Real Life Person pictures. User:Williama17 has this photo. But that means a more vivid background is introduced, and the AI has already a lot of faces to look at. Miis are simple, but the one thing that all miis lack are ears, so the AI will think that (alright guessing time) one out of 235 people have ears.

But what are you going to do about Favorite colors? Well the AI will be programmed to detect shirts. If a Mii is not wearing a shirt, the AI will either guess the Miis favorite color is the Mii's skin tone or the Mii has no visible favorite color. The AI will then have to guess the average Favorite color and say (mORE GUEssINg) "Okay, x% of miis like black, y% of miis like brown, z% of miis like light green, a% of miis like dark green, b% of miis like white, c% of miis like red, t% of miis like pink, n% of miis like purple, u% of miis like cyan, m% of miis like blue, r% of miis like yellow, s% of miis like orange, p% of miis like their skin tone, and f% of miis have no visible favorite color, and the AI will know that x+y+z+a+b+c+t+n+u+m+r+s+p=100. (The variables mean: X for example, Y for yield, skip z, A for any, B for base, C for count, T for this, N for number, U for unit, R for random, S for some, M for measure, and P for percent) so now the AI will have to place the Favorite Colors on its generated faces by the percentage of the favorite colors (so lets say X is 45%) so 45% of the 15 miis the AI generates (or 6 miis) like Black, where as 10% of miis (or 1 mii) likes brown and 0.1% of miis like their skin tone (or no miis) like their skin tone.

The AI will then start training, but since it hasn't seen anything, it just slaps a bunch of colors down and says "face!" but once it sees 8 photos (MORE GUESSING) (or 12.5% of the training set) it looks back at its work and says "not a face! i messed up somewhere!" but since there is 8 photos to 15 frames, the AI needs another set of 8 faces to fill in the frames. After the AI sees at least 45% of the data (G U E S S) (106 photos) it will know that random colors is not a face and that a face needs eyes, mouth, nose, and hair of any type (like eyebrows, scalp hair, facial hair, or body hair) and a skin tone. It also learns that (OH COME ON GUESS MORE) at least 62% of miis (146 miis) wear any form of glasses and 49% of miis 115 miis) have a beard or moustache or both.

Then the AI will see all the training data and sees like features it will use those features more often to create the miis.

The AI will then notice that each mii has two eyes therefore eliminating the eye placement.

if you have more questions ask me pls i will explain it in an update