[CuriouslyC]

I'm disappointed in Ted here. For a writer that likes to delve into the possibilities of tech, he's sharing a surprisingly underbaked view. People outside of tech seem to think that AI creation is just "fat finger a prompt -> take output and claim to be an artist on the interwebs" but the reality is that all the people I know who actually call themselves AI artists do photobashing, image2image, controlnets, inpainting, custom models, etc. Likewise, the people I know using AI to write fiction are meticulously developing characters, timelines, scenes, story arcs, style samples, etc and using AI to handle creating rough drafts that they then hand tune.

[ben_w]

I believe he demonstrated awareness of the difference between lazy use and effortful use; he appeared to me to acknowledge the latter as art:

"""The film director Bennett Miller has used dall-e 2 to generate some very striking images that have been exhibited at the Gagosian gallery; to create them, he crafted detailed text prompts and then instructed dall-e to revise and manipulate the generated images again and again. He generated more than a hundred thousand images to arrive at the twenty images in the exhibit."""

[becquerel]

Something I've had in the back of my mind is that gen AI has enabled a new generation of outsider artists. That's all. It has lowered the barriers to entry for creativity so much that a whole host of people who have had no formal training or dialogue with Real Artists are able to jump in and just make things they want to see. No surprise their creations are ugly or bad or soulless or weird by conventional standards; that's the norm for outsider art

[ToucanLoucan]

> People outside of tech seem to think that AI creation is just "fat finger a prompt -> take output and claim to be an artist on the interwebs"

Probably because that's the predominant experience of people encountering AI art on the Internet. I have no doubt whatsoever that there are people out there using AI to do interesting things, but like with basically every technology, the vast, vast majority of the output you're going to see is people who see a labor saving device that can make doing... something, at scale, brain-dead easy. Be that generating shitty coloring books and selling them to overworked parents, generating shitty books on niche, dumb topics like the crystal healing woo shit and selling them to uncritical audiences, or just generating page upon page of boring, shitty artwork and uploading it to DeviantArt and paywalling it.

And that's just individuals. Many online businesses are actively enshittifying themselves too, adding AI generated content alongside (or in place of) human created content. On the note of DeviantArt, they built an AI generator into the damn site so people can fill it with even more low-effort garbage than was already getting uploaded. And of course Google now headlines your search results with a shitty LLM summary that runs the gamut between "dull, uninteresting summary of somewhat relevant information" to "complete nonsense that actively endangers lives" while also depriving even more websites of even more traffic that gave them whatever information in the first place.

Like, again, I have no problem envisioning some people and some orgs some place are doing interesting stuff with this tech. However I cannot overemphasize how utterly, completely, totally dog-shit my experience personally has been with it and how harshly I now tend to judge any project parading around AI integration. I'm open to being wrong... but I'm usually not.

There was that Vaudeville game that made the rounds that I felt was at least trying to do something interesting with LLMs, but like... the tech just wasn't there yet. You're talking to characters and can say basically whatever you want, and then an LLM generates an answer based on the context of that character and it's read back to you by text-to-speech. It's... neat? For like ten minutes, and then you're just playing a detective game with impressively bad writing and zero-effort VO, and the fact that the entire game was built of pre-built, unchanged assets made it feel incredibly cheap and low-effort. The only thing it's really good for is as streamer fodder, weird garbage for people to overreact to and fuck with for an audience.

[Ukv]

I think this is mostly just Sturgeon's law (ninety percent of everything is crap) opposed to anything unique to ML tools. The vast majority of photographs you see on the Internet have also likely been quick phone-camera snaps, with comparatively very little "high art" photography. Lower entry barriers result in more works total, but with disproportionately many at the lower end.

There is still a lot of interesting work making use of ML tools. Maybe I'm biased towards art that embraces experimentation and new technology, but I found even images like https://i.imgur.com/Jybvj0r.png (zoom out) far more interesting than most of what I see in galleries.

[ToucanLoucan]

> I think this is mostly just Sturgeon's law (ninety percent of everything is crap) opposed to anything unique to ML tools.

I think the unique point to ML tools is the sheer volume that, IMHO, vastly outpaces 90%. And I think that's partially down to the fact that image generators are themselves, by their existence, low-effort easy-to-use tools to create images. It's not as though there wasn't already a vast and comprehensive existing tool-set for aspiring artists, even ones with not a penny to their name, to use. Tons of open source art programs exist, and if you are ready to jump to paying for your tools, you have an incredibly diverse set of options over all manner of capabilities, price points, and focuses. The notion that these tools "democratize" art has been silly to me from the beginning; there were already tons of tools available to anyone who wanted to learn the skills to use them. These tools, instead, seem directly aimed at people who don't want to learn those skills, and like... unskilled artisans don't make good things. Sorry not sorry. If you lack the interest in the subject to learn the skills of how to make a thing, you probably also lack the interest to learn what constitutes a good version of that thing, and even if your AI is very well made, you won't know what to really ask it for. Which I think is the reason behind so many prompts including stuff like "octane render, unreal engine, featured on artstation."

So while yes I agree in principle that Sturgeon's law is definitely at work here, I think it's important to note that the tools themselves are largely just... not really going to fit into a creative's workflow who has the skills already, not even just the literal skill to put a pencil to paper, but the skill to know what a good version of whatever they want would look like. It's the same reason I don't really use copilot all that much, because it's easier to just write the stuff I know needs writing than asking it to generate it, and then modify it to suit my code-style and existing environment. I don't find that a compelling time saver, it's more like a time cul-de-sac. Yes I'm spending the time writing the prompt instead of writing the code, but I'm frankly pleased as punch to just write the code.

I guess to TLDR my own comment here: if you knew how to make the things, you'd just make the things. Image generators are explicitly for people who don't know how, and that reflects in the quality of what's made.

[Ukv]

> And I think that's partially down to the fact that image generators are themselves, by their existence, low-effort easy-to-use tools to create images

Is the same not true of point-and-shoot photography? Or crayons? There's a near-endless supply of low-effort content due to tools designed to be easy-to-use. Anecdotally I still see more "crappy photographs" (many of which my own) than "crappy AI art".

With both you can get deeper into the details, making choices about ControlNets, LoRAs, inpainting, etc.

> [...] people who don't want to learn those skills, and like... unskilled artisans don't make good things. Sorry not sorry. If you lack the interest in the subject to learn the skills of how to make a thing [...]

I probably wouldn't make claims about ML tools "democratizing art", but at the same time I feel this is too reductive in the opposite direction.

There are reasons why working-class people are vastly under-represented in arts. I think limited ability to dedicate a huge chunk of time to a creative pursuit is a largely overlooked reason, not just lack of interest.

I think it's also fine to want to, say, design a game without hand-painting all the normal maps - instead generating them with ML tools based on your textures. Someone not specializing to have fine-level technical skills in all relevant areas doesn't imply lack of creativity/interest at a broader scale.

> I think it's important to note that the tools themselves are largely just... not really going to fit into a creative's workflow who has the skills already

I'd claim lot of ML tools, like generative fill built into various image editors, already do even for those who aren't going out of their way to experiment with ML.

Sometimes it's useful to work at the higher level allowed by automation, and sometimes it's useful to have fine-grained creative control. These aren't mutually exclusive - the approaches can/should be mixed where appropriate. I've had good success with sketching out an initial block-color image, then iteratively diffusing and tweaking it.

[ToucanLoucan]

> Is the same not true of point-and-shoot photography? Or crayons?

Those have a skill floor though, even if it is quite, quite low. If you can't manage to get the object you're trying to take a photo of in-frame, or manage to draw the thing you're trying to draw, there's no amount the tool can do to compensate for that.

> There's a near-endless supply of low-effort content due to tools designed to be easy-to-use. Anecdotally I still see more "crappy photographs" (many of which my own) than "crappy AI art".

I mean it depends how you define crappy photographs. My phone camera is a tool, and I use that tool to document things for all manner of purposes. I wouldn't call those photos artistic in any way at all. It feels like you're deliberately saying "all photos are art, and most of them are bad" when I think the vast, vast, vast majority of those, including by the people who took them, would not be referred to as art.

> There are reasons why working-class people are vastly under-represented in arts. I think limited ability to dedicate a huge chunk of time to a creative pursuit is a largely overlooked reason, not just lack of interest.

Agreed wholeheartedly. But, a working class person who has things they want to express artistically is going hit various walls with generative models very quickly, in much the same way I did. Like, if you feel a creative verve at all, I just can't fathom you looking at the wide assortment of all manner of tooling, and choosing the one where you're playing telephone with a toddler that paints over-smoothed, nonsensical photo-realistic pictures.

And again we go back to the notion that "the process is the point" and as a creative, I completely agree. There are certainly times I feel frustration at my tools and wish they would just make what the hell I'm trying to make, but if that was the entire process, I would get nothing from it. Figuring out what prompt will get you what kind of output is interesting, but it's not fulfilling.

> I think it's also fine to want to, say, design a game without hand-painting all the normal maps - instead generating them with ML tools based on your textures.

To be totally real I've never heard of someone drawing normal maps. I thought the traditional way you went about making those was having a high-detail model inside a low-detail one, and generating them that way.

> Someone not specializing to have fine-level technical skills in all relevant areas doesn't imply lack of creativity/interest at a broader scale.

It's not a matter of high or low skills, it's a matter of wanting skills and wanting easily made repetitive crap. If you're the kind of person who finds it fulfilling to slam text into one of these things and get your teddy bear smoking weed pictures, and that's all you want and are fulfilled, more power to you. I wouldn't personally call that art, nor would I find it nourishing to my creative spirit, I would say that's just instant gratification and there's absolutely nothing wrong with that. Now if you take that stuff and then go try to sell it... I mean that's your prerogative, I'm definitely not buying and I would encourage anyone else to just type a similar prompt into a generator and get it that way.

[Ukv]

> Those have a skill floor though

Point-and-shoot cameras, finger painting, or crayons have a lower skill floor than even basic text-to-image generation, I'd claim. You can give those to children prior to the age where they'd have a proper grasp on describing visuals through language/writing.

Yet, I don't feel as though the glut of low-skill content subtracts from any of those mediums - regardless of whether you disqualify a child's macaroni art from being art. Probably even the opposite; I've enjoyed areas that have lowered the technical skill barrier to allow people to create who otherwise wouldn't have been able to (like the creative explosion around Flash games, with ActionScript and the tooling being relatively beginner-friendly) in addition to it leading to more (even if proportionally less) high-skill content.

> But, a working class person who has things they want to express artistically is going hit various walls with generative models very quickly, in much the same way I did. Like, if you feel a creative verve at all, I just can't fathom you looking at the wide assortment of all manner of tooling, and choosing the one where you're playing telephone with a toddler that paints over-smoothed, nonsensical photo-realistic pictures.

I think the walls of what's possible using generative techniques in your workflow are almost by definition* further out than with only traditional techniques, and that the idea generative tools must be like "playing telephone with a toddler" comes largely from not having tried out most of the generative tools available or typical workflows.

I'd recommend checking out ComfyUI, starting with some existing examples (https://comfyworkflows.com/ seems to show workflows, when you click on the image) then playing around to see what's possible. Or for something a bit less technical, NVIDIA Canvas is fun, and useful for skyboxes: https://www.nvidia.com/en-gb/studio/canvas/

*: For a while 3D ML tools in particular did somewhat lock you out of other tools due to working on NeRF representations, but increasingly there's the option for regular meshes with sensible topology.

> And again we go back to the notion that "the process is the point" and as a creative, I completely agree. There are certainly times I feel frustration at my tools and wish they would just make what the hell I'm trying to make, but if that was the entire process, I would get nothing from it. Figuring out what prompt will get you what kind of output is interesting, but it's not fulfilling.

Do you not think you could be fulfilled with tools that let you focus on the bigger picture? I've worked with traditional procedural generation for cityscapes before and I don't feel it necessarily took away - just widened the scale I could create at, while still allowing me to zoom in and tweak individual buildings where I needed to.

> To be totally real I've never heard of someone drawing normal maps. I thought the traditional way you went about making those was having a high-detail model inside a low-detail one, and generating them that way.

If you have a 3D mesh sculpted then yeah you'd bake its normals from geometry - but you don't have that if you've just, for instance, painted some planks texture in Photoshop. You could hand-paint a normal map, hand-paint a height map and generate the normal map from that, or - as is increasingly common - generate normal/specular/roughness based on texture.