Hello everyone!

I've been wondering on the internet and I came across this: https://everypixel.com/aesthetics
Seeing what this can do I tought that we could use this on this site. For example it seems a neutarl could tag bacground color. Because it wouldn't miss tag it for sure.
Sadly I don't know much about neural network programming, but I think this would come quite handy in tagging, if ye could make a working system somehow.
What do you guys think?

Hexdragon said:

Hello everyone!

I've been wondering on the internet and I came across this: https://everypixel.com/aesthetics
Seeing what this can do I tought that we could use this on this site. For example it seems a neutarl could tag bacground color. Because it wouldn't miss tag it for sure.
Sadly I don't know much about neural network programming, but I think this would come quite handy in tagging, if ye could make a working system somehow.
What do you guys think?

May have an unintended side-effect of TWYK if it relies on information from other posts.

my avatar (post #928980)
2.0% chance that this image is awesome
Sport Equipment Black_Color Power No_People

kamimatsu's avatar (post #914278)
99.3% awesome
Illustration Human_Face Fantasy One_Person People Halloween Spooky Cultures Men

Interesting that it applied the "Men" tag to a girly male. It seemed to detect solo as "One Person", but also added "People".

Hexdragon's avatar (post #708221)
17.9% awesome
Animal Computer_Graphic Horse Vector Clip_Art Image Mammal Mythology Monster-Fictional_Character Backgrounds Dragon Carnivore Nature Wildlife Illustration Fantasy Fairy_Tale Green_Color Characters

It detected "Dragon" but thought it was a horse too. "Carnivore" might be based on the teeth and tongue_out.

Hexdragon said:
For example it seems a neutarl could tag bacground color.

Detecting things like purple_theme is not too hard. @Genjar made an extension that could do it.

The awesome score is probably meaningless to us. Some tags are applied correctly on accident but with low confidence (dragon and horse, one person and people). A version trained on e621's images and the tags used on the site could do much better.

A machine learning tagger will have a place in e621's future but it could be a difficult and expensive project. The longer we wait, the better the software and hardware gets and the easier it will be.

post #1174553
GOOD: Animal, Happiness, Smiling, Cheerful, Vector, Illustration, Cartoon, Cute, Drawing - Art Product, Illustrations and Vector Art, Computer Graphic
MEH: Humor, Backgrounds (abstract_background), Design (abstract_background), Symbol (the carrots?)
DUMB: Characters, Mascot, Nature

post #1174563
GOOD: Human Face (interesting because some of the goof_troop characters look very human), Women, Decoration (giving it because of hair flower)
MEH: Backgrounds, Fashion, Red (it needs to be more specific to be useful, in this case red_hair)

post #1174577
GOOD: Men, Sky, Nature, Cloud - Sky, Outdoors, Sunset, Animal, Sun, Vector, Illustration, Computer Graphic, Cartoon
MEH: Backgrounds, Night (debatable, it's twilight), Blue
DUMB: People (do I interpret this as duo/group, human, or anthro?)

post #1174533
GOOD: Women, Females (here we see both Women and Females for some reason), Beautiful (subjective but what the hell), Human Hair, Beauty, Sensuality (warranted given the subject matter, but subjective), Cute (this tag is on the post), Young Adult (actually useful), One Person, Human Face (debatable), Glamour (virgin_killer_sweater), Teenage Girls, Elegance, The Human Body (debatable), Vector, Illustration, Fashion (too broad to mean anything?), Computer Graphic
MEH: Backgrounds, People (what does it mean?)

post #1174543 dickgirl, the greatest test of all
GOOD: Beautiful (again, subjective), One Person, Adult, Beauty, Fashion Model, Human Hair, The Human Body, Cute, Elegance, Illustration, Vector, Fashion, Style, Isolated (synonym for solo?) Cartoon
MEH: Retro Styled (you be the judge), Old-fashioned (you be the judge), People
DUMB: Women, Females

post #1174540
GOOD: Pattern, Design, Flower, Abstract (it is an abstract_background, tag it plz), Elegance, Pink Color (call it pink_theme), Cute, Animal (feral), Vector, Illustration, Computer Graphic, Illustrations and Vector Art
MEH: Backgrounds, Cartoon
DUMB: Decoration (background, not decoration), Fashion (she's nude), Nature

post #1174541
GOOD: Women, Females, Beauty, Beautiful, Sensuality (that's clothing for the bedroom), Fashion, Elegance, Adult, Human Hair, The Human Body, Isolated, Love, Glamour, Vector, Illustration
MEH: People, Art (lol), White (monochrome or purple_and_white would be better)
DUMB: Backgrounds, Young Adult (looks mature instead)

post #1137927 "What?" said the neural network.
GOOD: Characters (?), Humor, Vector, Cartoon, Illustration, Computer Graphic
MEH: Cute, Fun
DUMB: Backgrounds, Men, People (?), Education, Clip Art, Blackboard

post #1173821 (81% chance awesome? more like 100)
GOOD: Women, Nature, Beautiful, Beauty, People, Outdoors, Females, Caucasian Ethnicity, Love, Young Adult, Human Face, Adult, Sensuality
MEH: Cute
DUMB: Animal (no ferals)

The male was not detected.

What have we learned so far?

• We can see that it doesn't understand the concept of anthropomorphic animals, but it makes an attempt to by adding "Human Face" or "The Human Body" on anthro images, or "Animal" on feral images.
• The developers hand-picked a set of keywords. So when you see "Women" or "Teenage Girls", it is not necessarily trying to say it found a plural amount of females. "People" could be the broadest tag of all, applying to solo anthros despite them not being human. "People" probably means "one or more humanoids". So it would apply everywhere except on feral pics.
• It seems good at detecting emotions, smiling and subjective stuff like sensuality.
• It does not handle dickgirls well.
• The concept of sex is boiled down to sexy, aka "Sensuality".
• It is good at detecting outdoors, the sun.
• It has a mixed record on abstract backgrounds.
• It took fucking forever to test because I can't copy and paste the tags.

post #316778

animal computer_graphic vector image nature sea clip_art blue underwater smiling backgrounds water cute wildlife cheerful

4/10. Higher because I like that it picked up cute and cheerful.

And for fun

post #6268

vector illustration animal nature cartoon bird backgrounds drawing_-_art_product wildlife computer_graphic mammal tree illustrations_and_vector_art cute animals_in_the_wild branch

76.0% chance this image is awesome

cute

hehe

Lance_Armstrong said:

• It took fucking forever to test because I can't copy and paste the tags.

^^^

To copy text from the page, View > Page style > No style
(in firefox, not sure about chrome)

post #768534
Backgrounds, Illustration, Spirituality, People, Religion

Well, that was terrible. Or maybe I'm just biased against it because it keeps saying that all of my photos are 0.0% awesome. Except the ones that nobody likes.

Ijerk said:
To copy text from the page, View > Page style > No style
(in firefox, not sure about chrome)

Thanks. Here's some more.

Genjar said:
post #768534
Backgrounds, Illustration, Spirituality, People, Religion

Well, that was terrible. Or maybe I'm just biased against it because it keeps saying that all of my photos are 0.0% awesome. Except the ones that nobody likes.

I do get a Raptor Jesus vibe from around the head.

my avatar...

5.3%
chance that this image is awesome

illustration
vector
cartoon
humor
animal
happiness
design
fun
flying
cheerful
bird
computer_graphic

nobody's fed it the cheese grater pic yet

Munkelzahn said:
nobody's fed it the cheese grater pic yet

Scroll up.

The neural network found my avatar pretty awesome.

It doesn't work on my browser for some reason, so I can't do any tests.

One main issue is that the lifespan of a neural network is directly tied to the space it has. As it runs out, it needs to unlearn things, meaning it has to assign space for what to unlearn. It makes increasingly poor decisions on what to erase until it becomes unusable. Basically it gets the AI equivalent of Dementia.

post #1164592
95.9% chance that this image is awesome
Vector, Computer Graphic, Animal, Cartoon, Clip Art, Men, Characters, Cute, Smiling, Mascot, Image, Illustration, Backgrounds, Humor, Mammal, Blue, Animated Cartoon, Happiness, Small, People

Very interesting set of tags.

kamimatsu said:
One main issue is that the lifespan of a neural network is directly tied to the space it has. As it runs out, it needs to unlearn things, meaning it has to assign space for what to unlearn. It makes increasingly poor decisions on what to erase until it becomes unusable. Basically it gets the AI equivalent of Dementia.

Interesting. Sure, humans forget things but I once learned that the brain basically doesn't run out of memory space. It creates more if needed.

kamimatsu said:
One main issue is that the lifespan of a neural network is directly tied to the space it has. As it runs out, it needs to unlearn things, meaning it has to assign space for what to unlearn. It makes increasingly poor decisions on what to erase until it becomes unusable. Basically it gets the AI equivalent of Dementia.

Storage is still getting cheaper and denser.

http://www.zdnet.com/article/samsung-crams-32tb-ssd-into-a-2-5-inch-drive-aims-for-100tb-by-2020/

post #1162438

98.7% chance that this image is awesome.

I can agree with that.

Sorrowless said:
Interesting. Sure, humans forget things but I once learned that the brain basically doesn't run out of memory space. It creates more if needed.

You learned correctly. In fact, as long as someone can take in new information, it slows the rate of Dementia, and pieces can adapt to learn tasks they weren't initially for. Basically, for the human brain, more information means higher capacity.

Accidentally tagging something wrong is already easy for a human. A machine that can't see the context of any visuals it scans would be even worse in that regard.
Especially if it automatically adds the tags.
And even if it just feeds suggestions, we already have the related tags tool, which is much better for contextual tag suggestions that are based on the tags already present.

Quasi said:
Accidentally tagging something wrong is already easy for a human. A machine that can't see the context of any visuals it scans would be even worse in that regard.
Especially if it automatically adds the tags.
And even if it just feeds suggestions, we already have the related tags tool, which is much better for contextual tag suggestions that are based on the tags already present.

It should be easy to prove how well a machine tagger would work before using it on the scale of every image uploaded.

That's why I ran those tests above. To see how well it would work. It did not work very well.

A neural tagger trained on e621's own images and tags would probably do better. And the technology will get better, because Google and other companies are working on image tagging. Not only that, they want to tag parts of the image which could help for combined tag searching (example: finding human males instead of images with a human female and an anthro male in it).

Quasi said:
Accidentally tagging something wrong is already easy for a human. A machine that can't see the context of any visuals it scans would be even worse in that regard.
Especially if it automatically adds the tags.
And even if it just feeds suggestions, we already have the related tags tool, which is much better for contextual tag suggestions that are based on the tags already present.

Ignoring context is actually a kind of good thing. We look at the raw imagery.

Sure, it's not perfect, there are points where context is allowed, but people might not even think of some of the things it can suggest.