• DavidGarcia
    link
    fedilink
    English
    arrow-up
    2
    ·
    28 days ago

    It’s not so much that I want to automate everything, but that I want to only have to manually do what I like doing. I’m very picky and if I go through e.g. ImaginaryNetwork on Reddit, 99 of 100 images there aren’t interesting to me. It’s wasting 99% of my time. I’d rather prefilter it and only look at the 5-10 images that I’d probably like.

    Also, I’m trying to cover every category of “stuff” with high quality images in my collection. Mainly so I can use them for reference. By category I mean characters, animals, environment, tech, builds, art style etc…

    Thinking up some story pitch and then filling out the blanks with good art is my idea of fun. Maybe I need some sort of concept art of some creature. So then I need to be able to only look at all images of creatures, instead of going over 30k images, 95% of which are irrelevant to the task.

    That’s why I need some sophisticated tagging system. If I had to do that manually I would literally have to write 300k tags or something. It’s sadly unrealistic for me to go over 30k images over and over again, but I still want to use them somehow.

    Another issue is the distribution of categories in online communities 99% is hot girl art, not that I don’t enjoy that too, but its irrelevant when you are looking to fill out all “categories” of art. Some categories are just completely overrepresented. It would be cool if I could specify 20 categories and then only look at an equal number of images from each categories. Otherwise I’m stuck in the 1 millionth reursive loop of downloading images of run down mediterranian looking masonry or something. 90% of my library is characters, but I also really like tech art, but there just isn’t that much of it out there in comparison.

    Some day I might want only enjoy looking at images of category A instead of category B, but all my usual channels are flooded with B. It would be cool to have something to only look at A.

    I only really have a couple of hours a week where I can really enjoy browsing art, so I’d rather make the best of it.

    If I can save time on finding new or relevant images, I can spend more time doing what I actually enjoy most.

    I’m so tired of endlessly compulsively scrolling, because I need to find just one more specific image in a sea of irrelevant images.

    I’m thinking my library should be big enough now to serve as a reasonably good approximation for my tastes, so a good model should be able to only give me interesting images.

    Of course you have to play around with how similar the images you filter in are supposed to be. But you can do extremely cool things with embeddings you couldn’t ever do with words.

    Like extract the essence of a very specific art style or composition and look for similar images. For example you could average the embeddings of a certain category and extract what they have in common only. I would be interested what images I would find if I used the average of all embeddings of my entire library as a filter.

    You could uniformly sample from some latent space. Or you could even look for the most dissimilar images to your library so you can discover something new.

    So many cool applications for embeddings.

    None of the existing image similarity search tools that exist are specific enough for my tastes. I always end up looking at 90% irrelevant stuff again.

    Hope that makes sense.