[HN Gopher] Artist uses AI to fake 70s science fiction pulp cove...
___________________________________________________________________
Artist uses AI to fake 70s science fiction pulp covers - artwork
and titles
Author : glitcher
Score : 80 points
Date : 2022-02-07 19:37 UTC (3 hours ago)
(HTM) web link (cdm.link)
(TXT) w3m dump (cdm.link)
| JoeAltmaier wrote:
| My son put up a similar AI portal for his mother (who is a writer
| and artist). It takes a title and creates the art. For instance,
| I entered "Ford F150 by Frida Kahlo" and got a pastiche of red
| truck parts, faces and the letter 'F' followed by three or four
| quasi-symbols. All in a saturated style similar to Kahlo's
| paintings.
|
| I'm guessing the faces because she did so much portraiture.
| charcircuit wrote:
| It doesn't seem to be coherent in what it wants. As you go down
| in the image things change into different things.
| Laylo_ wrote:
| Apparently orange circles were big in the 70s (or at least in the
| training dataset)
| Lammy wrote:
| My headcanon is that everything was brown and orange in the
| '70s because it won't show nicotine stains as easily.
| rightbyte wrote:
| I find these kind of art kinda annoying to look at, because it
| takes like 3 or 4 secs until you realize that you are watching
| nonsense. The initial feeling fools me.
| jcims wrote:
| >The initial feeling fools me.
|
| I'm the exact same way and I think this says something about
| how the methods used to generate these images relates in some
| way to how our mind processes them.
| visarga wrote:
| Seems similar to the incoherent material in doodling, role
| playing, or dreaming. Humans do a lot of that.
| Cerium wrote:
| I love it. Unlike many machine learning examples this one does
| not highlight how good the model is, but highlights how poor the
| source material often is. Many of these are completely believable
| covers.
| jhbadger wrote:
| Indeed. I have a soft spot for 1960s-1970s SF paperbacks and
| always end up with a few when I visit used book stores. These
| look uncannily like the covers of those.
| TheOtherHobbes wrote:
| There was nothing poor about the source material.
|
| This actually highlights how this kind of AI simply smushes
| sources up randomly without any sense of content or meaning.
|
| It has a look - more horror than SF - but that's because
| distorted smushing with vague hints of distorted body parts is
| a look in itself. (See also artists like Francis Bacon.)
|
| Neither AI nor developers aimed for this look _deliberately_.
| It just happens to be what falls out of the process. (See also
| Wombo, which has a similar aesthetic.)
|
| What it doesn't do is evoke the imaginative narratives of the
| originals, which were consciously crafted to suggest a mood,
| setting, or context. Compare with (say) Bruce Pennington or
| Chris Foss and it's easy to see what's missing.
| TaylorAlexander wrote:
| Thanks for mentioning some original artists. One thing I am
| realizing here is that the original artists who worked hard
| to create the source material for these algorithms are often
| totally overlooked. "I threw a bunch of book covers in to an
| algorithm and look at all this cool output!" Sure, but who
| are the people who created all that source material? Can we
| credit their names as we build off of their work? We can and
| should credit the algorithm developers too and the person who
| put all of it together. But I scanned through TFA and I don't
| see the names of any original artists or mention of the
| source material, and that seems like quite an unfortunate
| oversight.
| Lammy wrote:
| > The Past is Fictional
|
| It's interesting to think about how we tend to grade the "smarts"
| of machine intelligence by comparing them to what us humans
| believe to be the truth. Even if they were trying to teach us
| something how would we ever know?
| JoeAltmaier wrote:
| SMBC: https://www.smbc-comics.com/comic/gan
| pessimizer wrote:
| These covers aren't great, but if you could design a process that
| could produce great ones, then write books to fulfill the best
| ones, you wouldn't be that far off from the editorial process of
| a lot of pulp publishers that wrote books based on good covers.
|
| AI generated covers and titles used as writing prompts would be
| very fun. They would have to be far better than these, though.
| all2 wrote:
| Soooo...
|
| There's this? https://tvtropes.org/pmwiki/storygen.php
|
| And that is a wonderful way to get ideas. :D
| mcbuilder wrote:
| One of the more refined and well executed AI art projects I've
| seen! The subject matter has a nice eerie connection.
| bryanrasmussen wrote:
| This might have been useful to me in early eighties high school,
| I had a class in which I had to write book reports, once per week
| I think, and there was a sort of template as to how they had to
| be written with analysis of character motivations etc.
|
| It's been a long time but I found something about the critical
| methodology used just extremely insulting to any piece of
| literature it was applied to, reductive and harmful to actually
| appreciating anything, so despite reading a minimum of 4 books a
| week I decided to write fake book reports. Generally crappy
| science fiction and fantasy was the genre.
|
| I only remember one fake book I wrote a review of - The
| Hinterlands of Horlon - a crappy Tolkien rip off with a quest and
| everything, written by James R. Canton which was also a role
| playing character name I had for some occult spy game I was in at
| the time. Fake cover art would have been a nice topper to the
| fraud.
| pointlessone wrote:
| I feel like this technique would be much more successful for body
| horror genre.
| candlemas wrote:
| All of the AI art I see has a sort of melted look.
| iakov wrote:
| I agree. It's a mish-mash of melting shapes, nothing resembles
| anything.
|
| I'm really confused why would someone call this "perfectly
| fake". Have those people seen the real covers? The covers were
| meant to catch your eye, to evoke some kind of emotion.
| Compared to them, most of the examples evokes only repulsion in
| me.
| JoeAltmaier wrote:
| I have the same reaction. They're like an alien with only a
| description of human activities but no real knowledge of
| behavior or society, tried to create a painting. Everything
| is strangely squished, there's no clear dividing line between
| one thing and another, nothing is a functional symbol but
| just an impression of a symbol. "Humans like yellow; lets put
| a yellow shape in there!"
| amelius wrote:
| They should put a human in a fMRI scanner, and place this in
| the feedback loop of the GAN. When the human feels repulsion,
| the fMRI scanner picks this up and sends negative feedback to
| the network.
| octopoc wrote:
| The human would most likely quickly be desensitized and
| only output feelings of self hatred and suicide. But to be
| sure we should start a multibillion dollar startup that
| generates jobs for millions of starving third worlders.
| tombert wrote:
| There's a part of me that actually wants to write a short story
| called "Green Glass is the Color of the Wind" now, since I think
| that's a pretty funny title.
| fckthisguy wrote:
| "The moon is just sunlight" was my favorite. I'd read that!
| AdmiralAsshat wrote:
| What we need now is a second AI to generate book review blurbs
| to put on the back cover.
|
| "Wondrous...thought-provoking...panda." - The NY Times
| kwertyoowiyop wrote:
| How long until GPT-3 can do that?
| tombert wrote:
| That honestly might be more appropriate. A machine learning
| AI drawing a cover and a machine learning AI writing a book.
|
| The singularity might come sooner than I thought.
| visarga wrote:
| They can be the same AI, cross-modal training is working
| very well.
| jcims wrote:
| I haven't tried this yet but I'm sure you could coax GPT3
| into describing images as part of an illustrated book, then
| feed those into one of these generators.
|
| Welp, i just tried and: https://imgur.com/x9bwRq2
| JasonFruit wrote:
| "Bizarre pastiche" describes the training material as well as the
| output. Well played!
| marstall wrote:
| this artist has captured so many of the aesthetic qualities of
| these classic cover images here. One that is missing, I noticed,
| is the quality of sparking my imagination, or the evocation of a
| world. There's nothing in those images that is connecting with me
| on that kind of level.
|
| Which makes me wonder if this is a problem with many/most of the
| AI images one sees. I think so?
| visarga wrote:
| Neural nets can paint whatever you like, but you have to choose
| the right training material, model, prompt, etc.
| reaperducer wrote:
| So, who's the real artist here? The AI, or the guy who programmed
| the AI?
| Jack000 wrote:
| In photography it's neither the camera or the camera
| manufacturer that's considered the artist, I think the same
| logic would apply.
| themikesanto wrote:
| Yes!
| rectang wrote:
| It seems that the original artists who created the source works
| feeding the model did not even enter your consideration. :(
|
| Looking at these images I am all but certain that there are
| going to be identifiable portions of the originals in the AI-
| generated images. They are derivative works. Erasing the
| original artist's role in creating these images is immoral.
| cellis wrote:
| A human brain can only paint what it has observed or
| derivatives of observation. How is digital remixing any
| different?
| twox2 wrote:
| I love how the images include the wear and tear of old paperbacks
| giving these covers an authentic look... but I think the titles
| need some slight photoshopping to reflect those scratches :)
| bryanrasmussen wrote:
| I notice the only cover with an author popular enough to have
| their name bigger than the title of the book is One Telling Ten
| Thousand by Isabella V. Rivera which does seem to go against the
| whole boys club feel of Science Fiction of that time. I think it
| would have been more likely a gender neutral name like Andre
| Norton's, maybe Kim Rivera or V. Rivera.
| jcims wrote:
| It's pretty easy to dismiss this stuff but I've found it has been
| very fun to play around with and certainly has allowed me to
| create imagery that far exceeds my skill.
|
| Twitter seems to have an endless stream of examples with the
| aiart and generativeart hashtags. Also I really like Katherine
| Crowson's twitter because she builds Colab notebooks that n00bs
| like me can use to tinker:
|
| https://twitter.com/RiversHaveWings
|
| https://github.com/crowsonkb
|
| Some examples I shared a while back using Katherine's VQGAN+CLIP
| notebook: https://imgur.com/a/1R5UZQb Promp was 'mad max
| dieselpunk alien spaceship landed in a desert' and then I
| upscaled the images. These are cherry picked from a set of about
| 3x this size.
|
| One thing I find pretty remarkable is how good it handles shadows
| and depth of field. These are all kind of mud tones but if you
| let it run away with color you'll get some really brilliant
| results.
| devindotcom wrote:
| I was just looking at some of her stuff yesterday, very cool.
| I'm even more of a noob though, since I'm not sure how to get
| this code (similar to other notebooks I've played with) into
| Colab to test out. If there's an obvious first/next step I
| should be taking, could anyone point me in that direction?
| jcims wrote:
| I think you need a google account to access free collab
| runtimes. If you have one, click here - https://colab.researc
| h.google.com/drive/1QBsaDAZv8np29FPbvjf...
|
| Scroll down to prompts and put in your prompt. I just did:
| prompts = ['fire ants attacking a school bus cinematic 35mm']
|
| (re 'cinematic 35mm', you'll find that there are little
| bumper phrases, like 'in the style of <artist>' or 'trending
| on artstation' that sort of thematically guide the output)
|
| Click on 'Runtime->Run all' and let the notebook execute it's
| initialization steps and then start to generate the image. By
| default it goes 1000 iterations, and spits out what it's
| working with every 100 iterations. Once you get one you like
| you can update the prompt, re-run that cell and the one below
| it to generate something new.
|
| I pay $9/mo to get access to better GPUs (usually P100
| sometimes get better), the free accounts tend to get slower
| ones so it might take 30-60 minutes to complete per image.
| devindotcom wrote:
| Oh yes I pay for Colab - didn't see this linked from the
| other page, or rather wasn't sure where to find it in the
| various notebooks. This is exactly what I was looking for -
| setting up to do a run now. Thanks very much.
| jcims wrote:
| No problem. This is an older notebook. Poke around,
| Katherine has posted more recent ones and lots of people
| have added bells and whistles to make them easier to use.
| Have fun!
| nix23 wrote:
| Wow looks super bad...except the fonts. It's just pixelvomit.
| yeezyseezy wrote:
| There's something deeply unsettling to me about art like this,
| as if one of the hidden layers is "colonoscopy"
___________________________________________________________________
(page generated 2022-02-07 23:01 UTC)