Why don't indie devs use AI-generated images as art?

[thesecret1]

These "AIs" are just following patterns. What they output are usually patterns that have been repeated countless times across thousands of works. If you, as an artist, can create something outside these patterns, or surpass them in quality in some way, then you have a competitive advantage over the AI and need not fear it. If you cannot, then you are the equivalent of a factory worker, doing the same thing over and over. Factory jobs of this nature were replaced by machines in massive amounts, and we do not mourn that. Why should we mourn you?

The screeching is especially funny since they are right that it tends to lack soul (ie. that it often looks the same as millons of other generic images on that topic, because of course it does; that's how it works), yet fail to realize that their works lack it as well.

 

[Strig]

I've seen more than a few people spell out similar tosh. The simple fact of the matter is that well executed artistic efforts follow patterns, art is pleasing to the eye in large part due to composition and geometry, even if what is being shown is fairly nonsensical and impressionistic (and many of the AI images are just like that, with good composition, good basic shapes, nonsensical detail - of course it's getting better by the day). Modern artists create works "outside of these patterns" all the time as they require neither skill nor taste. And in many cases can be recreated by children or even monkeys. Again, ad nauseam, if you remove the opportunity for mediocre, formulaic artists to earn a living there will be almost no good or great artists in the future, because you eliminate the incentive for growth and there are very few savants who enter the scene at a master's level.

Factory jobs of this nature were replaced by machines in massive amounts, and we do not mourn that.

https://www.azquotes.com/quote/749479

For you :3

 

[Peachcurl]

In other news, Jeffrey Combs photo from Re-Animator + "gentleman in glasses":

Same, plus the word "grim":

I wonder where the Joker-esque influence comes from. He never played the Joker, right? I know he's been involved in various batman franchise products.

 

[infidel]

I wonder where the Joker-esque influence comes from. He never played the Joker, right? I know he's been involved in various batman franchise products.

No, I don't think so, his actor roles are rare cameos in mainstream productions. He did some great voices for DC animated series. I don't see anything resembling the Joker in these images, what are you talking about? And which specific Joker are we talking about?

The base image I took was a cut out of this shot:

 

[Peachcurl]

I wonder where the Joker-esque influence comes from. He never played the Joker, right? I know he's been involved in various batman franchise products.

No, I don't think so, his actor roles are rare cameos in mainstream productions. He did some great voices for DC animated series. I don't see anything resembling the Joker in these images, what are you talking about? And which specific Joker are we talking about?

I'm mainly talking about the top right picture in the "grim" set (though after looking at that I noticed at least some "lipstick" in the other set)

 

[Nazrim Eldrak]

I wonder where the Joker-esque influence comes from. He never played the Joker, right? I know he's been involved in various batman franchise products.

No, I don't think so, his actor roles are rare cameos in mainstream productions. He did some great voices for DC animated series. I don't see anything resembling the Joker in these images, what are you talking about? And which specific Joker are we talking about?

The base image I took was a cut out of this shot:

Perhaps the model interpreted/learned something from this scene.

I mean the pen looks the same.
His lips are red and the facial expression looks the same as at 1:48 min in the video
I understand this seems far fetched but you never know how the system will learn and interpret things I guess.

 

[infidel]

I'm mainly talking about the top right picture in the "grim" set (though after looking at that I noticed at least some "lipstick" in the other set)

I think this image is just full of bugs that happen when dealing with AI. That smile thing of his is made out of glasses temples if you look closely. The white lines below I think are just parts of smoke that bugged out.

Note how I used "gentleman in glasses" for tags since using just "gentleman" results in AI often picking images of people without glasses, then blending them resulting in errors like this:

I mean the pen looks the same.

I specifically removed the pen from from the image I used since at first the results were more like this:

The pen was treated as abstract "any prolonged object near the mouth I can find". Hehehe, good thing they don't allow porn . My layman's theory is that it looks for images that it thinks are "close" to what was supplied or described then layers them with some magic algorithm. In v3 the algo worked differently - it didn't cut out parts of the images well enough, so you get a distinctive MJv3 look that you can recognize easily:

Note how you literally cannot define what anything is except maybe the "floor" and "passage" and pieces of "stairs". After checking out what does it think about "cyberpunk city", I quickly zoned out:

Also it just couldn't do eyes without crazy glitches at all:

But MJv4 (I think) was taught to recognize the distinctive parts of image. Eyes are now eyes, faces are faces, etc, etc. After fiddling around, I get some sense of how it's working and what are the limitations of the starting images you need to supply if you want to get good results. If you want good faces, remove everything else except the face from the photo, otherwise if will result in bugs or lower pool of "close-looking images". If there's a cigarette or glasses, hint it with text. Try black and white image separately, since colors are also taken from the starting image as a "closeness" measure. Any images with large enough faces now looks pretty much perfectly:

Note that all of these are failures, btw. The goal was to make a "sasha grey vampire". And currently the AI studiously removes any and all vampire fangs even if they are on both of given images. "large fangs" does nothing.

Some things are still really hard for it. Yesterday I spent an hour or so trying to get a simple image of "woman in long dress in front of a car". And I conclude, that it is currently impossible to do that or my AI-query skills are not high enough:

Faces WILL be destroyed. And often figures. My guess is that it combies ladies from a set of images with cars from another set of images and that's where things get finicky really fast.

OTOH, would you look at that... Just a few of very specific words and...

Damn, where'd these one-eyed kids even come from??? All of these are going to be upscaled and then into my dark ambient music video once I finish polishing the track

 

[RobotSquirrel]

After checking out what does it think about "cyberpunk city", I quickly zoned out:

If you add more specifics it'll change up the style. Just setting it to Cyberpunk City gives you some pretty generic results.

This was with Blade Runner, Vivid, Lights, Billboards, Vehicles, H.R Giger. This was Stable Diffusion Automatic1111.
It really likes H.R Giger for some reason, I found the results were better than if you put in Syd Mead as the result is more 1960-1970s scifi which I guess is cool too but Giger gives you that clearly distinct 80s quality to it.
I also found the results were more pleasing in Portrait rather than in landscape, as landscape tends to have a lot of repetition, its a bit noisy. You get a lot nicer verticality on the skyscrapers in portrait and it makes the whole composition feel more boxed in and oppressive.

 

[Kev Inkline]

Automatic1111 is just the web ui, it doesn’t tell anything about the model that was used, but I suppose you used either as 1.4 or 1.5 base checkpoint.

 

[orcinator]

In other news, Jeffrey Combs photo from Re-Animator + "gentleman in glasses":

Been looking for an AIgen that modifies an existing image, can you spoonfeed me on which ones do this?

 

[infidel]

If you add more specifics it'll change up the style. Just setting it to Cyberpunk City gives you some pretty generic results.

Looks like MJv4 is starting to get ahead in this regard for the moment. I've only tried once (and FFS I'm staying out of that for the moment, I've been wasting all my time generating shit for the past five days or so). Image A:

plus image B:

resulted in C:

Very much reminds me of early 2000s videogame cover art for some reason.

Been looking for an AIgen that modifies an existing image, can you spoonfeed me on which ones do this?

I'm no expert, haven't even tried Stable Diffusion yet, but if you look at my previous posts, that's close to what I did most of the time with MidJourney v4. The caveat is that it's not "modify existing image" but "take image as a source and find something close to it if it exists with additional input, and if not, try real hard". My guess is that they're all do it that way unless you're using specifically AI filter apps.

EDIT: Here's a good one. My friends've been trying to do "lovecraftian girl" and "lovecraftian city" and stuff like that. I said it's as retarded as typing "lovecraftian tomatoes" because that either doesn't mean anything or is too vague, and I proceeded to try exactly that...

Spoiler

Well... Looks like I proved myself wrong.

 

[RobotSquirrel]

Automatic1111 is just the web ui, it doesn’t tell anything about the model that was used, but I suppose you used either as 1.4 or 1.5 base checkpoint.

Sorry I posted that at a really bad time. It was 1.4.

 

[Kev Inkline]

In other news, Jeffrey Combs photo from Re-Animator + "gentleman in glasses":

Been looking for an AIgen that modifies an existing image, can you spoonfeed me on which ones do this?

That's exactly what Stable Diffusion's img2img function does. You can adjust the deformation or noise from 0-100, where 0 means no modification and 100 means the original image is completely ignored and the image is generated 100% from your prompt only.

Check out this progression:

original

10% deformation or so

0.51

0.66

0.75

a detailed and atmospheric watercolor fantasy character concept art portrait of a lesbian librarian looking at the camera with an intelligent gaze by rebecca guay, simon bisley, Adriaen Isenbrant, larry elmore
Steps: 30, Sampler: Euler a, CFG scale: 8.5, Seed: 3262027661, Size: 512x512, Model hash: a9263745, Denoising strength: 0.51, Mask blur: 4

 

[infidel]

a detailed and atmospheric watercolor fantasy character concept art portrait of a lesbian librarian looking at the camera with an intelligent gaze by rebecca guay, simon bisley, Adriaen Isenbrant, larry elmore

Needs way more Biz, if you ask me but I like what I'm seeing (in the technical sense). Gonna try it at some point. How fast does it work and how heavy is it on CPU/GPU while doing it? Also where does it take the additional images it learns from? Does it come with a ready subset or do you need to train it yourself?

 

[Kev Inkline]

The base model is trained with several billion pictures, people have trained then their own models based on that. The current 'official model' of SD is 1.5, that has been trained on 595K pictures on top of 5.85B (Laion 5B dataset) its 1.2 checkpoint was trained on.

Here's a list of models, I don't know how up to date it is, and unfortunately many are hentai oriented or otherwise deranged.

https://rentry.org/sdmodels

Here's more details and resources:

https://www.reddit.com/r/StableDiffusion/comments/yknrjt/list_of_sd_tutorials_resources/

One way to have more nuanced and individualized models is to use embeddings, that are not computational as intense, but lend their definite own flavor to the images generated:

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion#training-an-embedding

 

[Kev Inkline]

Oh and you asked about the load on the gpu - I'm currently rendering these on my 2070S while waiting for the final parts of my rtx 3090ti build arriving, depending on steps and other details, it takes something around 5-10 sec per image att 512x512, more at higher resolutions.

 

[infidel]

Well, here it is

 

[V17]

Stable Diffusion 2.0 has been released and while it brings some interesting new features, it's also been nerfed in the NSFW department and as a result is worse at generating humans in general.

There's a lot of drama centered around that on Reddit because
a) The only usage they are creative enough to think of is "draw me a woman with pretty face and big boobs in the style of greg rutkowski" and since 2.0 is worse at that and no longer knows Greg Rutkowski (due to using their own CLIP model, this was afaik not intended and may be a temporary problem), they are seething, and because they are unable to think of any other usecase than their own, they also believe that SD is effectively dead now.
b) They are unable to comprehend that there is a difference between what is rational/ethical and what it possibly necessary in order to achieve a good outcome. Removing all NSFW imagery from the training dataset is probably unnecessary and won't make the world a better place in any way, but it's done to avoid the worst kiddie porn (or possibly deepfake) lawsuits, which could stop the open-source AI progress, and in that sense it might currently the least bad solution to a real problem.

The CEO promises that one of their efforts in the next months is to make training the AI easier and cheaper, including distributed training, so that one actor does not need to own a huge GPU farm, so it might be possible to train the porn back in. He also said that since the NSFW issue is now solved, updates will be released much more often.

I'm almost overcome with shadenfreude, because while I hate censorship, I fucking hate those stupid boring anime waifus and other dumb shit that floods SD communities and if I never have to see one more in my life I'll be happy. I do not understand how anyone spends time creating this garbage instead of something interesting or at least actual porn.

-----------------------

That said, an exciting feature of Stable Diffusion 2.0 is "depth2img" - you feed the AI an existing image, it tries to extract the depth data of the image (basically construct a 3D scene of it, understand the shape of the objects in the image) and then, according to your prompt, creates an image that has similar shapes as the one you provided.

This is exciting because it allows you to create things that are similar in shape but have completely different colors. Using img2img for this was difficult because you either got similar colors on the output or the AI strayed too far from the input image and created something that wasn't at all similar to the input.

So you can create a pose in a 3D software and tell the Ai to create a different character with that pose:

Some guy tried using it on sprite sheets - literally just downloaded a sprite sheet of a running man from the internet and told it to convert it to Thor, Spiderman and Trump. The overall quality is low because this is a vanilla model not trained on spritesheets, the images are very small and non-refined etc, but it's obvious that the process already works better than img2img and could be practical for videogames after finetuning. If you do this in img2img, the poses completely break.

 

[Kev Inkline]

 

[Dexter]

Spoiler

 

[Non-Edgy Gamer]

Stable Diffusion 2.0 has been released and while it brings some interesting new features, it's also been nerfed in the NSFW department and as a result is worse at generating humans in general.

Predictable. Every artist worth his salt has studied the human anatomy unclothed.

I told you all that that it would affect the model when they announced they were going this route.

 

[fuzz]

waiting for new roguelike tilesets to come out of all this.

 

[SlamDunk]

ArtStation's Artists Have United in Protest Against AI-Generated Images - 80 level

Thousands of creators demanded ArtStation to remove AI content from the website.

Over the past few hours, a new trend emerged on ArtStation among the platform's creators, who united in protest against AI-generated images in general and its presence on the ArtStation website in particular.

 

[rusty_shackleford]

ArtStation's Artists Have United in Protest Against AI-Generated Images - 80 level

Thousands of creators demanded ArtStation to remove AI content from the website.

Over the past few hours, a new trend emerged on ArtStation among the platform's creators, who united in protest against AI-generated images in general and its presence on the ArtStation website in particular.

how do they identify what is "AI art"?

 

[Non-Edgy Gamer]

how do they identify what is "AI art"?

What if they just removed all the formulaic, cheap art? (All of it.)

But seriously, someone will probably train an AI model for detection.

Didn't see these hypocrites do this for photobashing or tracing.

Now, AI art is going to be like steroids for athletes. With big name artists coming out and admitting they used AI art for parts of their piece. How laughable.