These look very nice, but are not pixel art in the sense I understand it. The pixel matrix is not consistent. They look pixelated, but if you zoom in, you'll notice that the block size varies quite a bit within one image and between images. There are also artifacts, pretty clear for example in the jumping deer's antlers.
I don't know what would be a practical way to do it, but I imagine some postprocessing step where a consistent matrix is enforced.
npinsker · 7h ago
Though you’re absolutely right, I feel like pixel art nowadays is more commonly thought of as just an aesthetic “feeling” — most indie games violate the rules of pixel art a bit by scaling and rotating sprites — and it doesn’t matter much to consumers as long as it looks good.
dwroberts · 7h ago
Scaling and rotating are normally runtime things the game does with the art though, and it’s a taste thing of ‘it looks bad’
Not adhering to a pixel grid is just… not pixel art.
mrob · 6h ago
I can think of one example of true pixel art that intentionally deviates from a strict pixel grid:
Look at the vegetation in the background. However, this is also a good example of the difference between real art and AI slop. There's human intention here, and I can guess a plausible scenario of how it happened.
I suspect the checkerboard pattern in the foreground came first. It has the practical utility of showing off the game's primary selling point, which is Sonic's high speed. You need a high-contrast pattern that's not too distracting and doesn't use too much memory. A checkerboard is an obvious choice. And once you've made that decision, it influences the aesthetics of everything else.
Checkerboards were a very popular texture in early 3D rendering. Any computer artist working in 1990 would be familiar with this, and would at least subconsciously think of 3D graphics. This is likely the reason for the stylized leaves on the trees, which look like low-polygon 3D models. In 1991 it felt futuristic. I believe the blocky look of the background vegetation is intended to convey this same "computer graphics" feeling. Drawing polygons is impossible, but drawing attention to the pixels gives the same impression of something futuristic.
This is the kind of non-obvious artistic decision that I don't believe current AI is capable of. A human intentionally reduced the graphics quality, and that actually made it look more futuristic. It's the kind of thing you see all the time in good art. Close examination reveals detail that deepens your appreciation for it. AI slop only imitates the surface polish of good art without including this deeper meaning. The closer you examine it the worse it looks.
joegibbs · 6h ago
What if it just gets scaled down with nearest neighbour? Then the pixels would be forced on to a grid of a particular size. Maybe some kind of algorithm that compares the % difference to the full scale image to see what scale makes the most sense for it?
Kiro · 3h ago
Yes, that works for most of the images.
kubb · 7h ago
It's also not animated, and there's not enough assets to build environments.
Don't get me wrong, some of these images do look good, they're just not enough for a game.
gametorch · 6h ago
Creator here.
You can create animated spritesheets with the click of a button. [1]
You can create as many assets as you want to make your game. It's all prompt-driven.
If you have ideas for things you want generated, let me know and I'll run your prompt for you!
Can you think of a single reason wy a game studio wouldn’t use this animation?
What are you doing to make sure your product meets the quality standards of the industry?
gametorch · 1h ago
Between the first commit for this project and today, which is a span of only 2 months, image models have increased in quality by an order of magnitude.
It's only a matter of time until the model is sufficient for everyone's standards.
The most puzzling thing to me is that in a forum full of technologists, a significant minority are betting against the advancement of technology.
If you could let me know why you seem so bent on my failure, I would be forever grateful.
Kiro · 7h ago
It's enough to just nearest neighbor scale it to the pixel grid to remove both artifacts and inconsistent pixel sizes. They are so few so it doesn't affect the final result.
mrob · 7h ago
Some of these have both highly variable block size and fine details, e.g.:
That's with today's models. Six months from now they will be exponentially better and cheaper.
Kiro · 4h ago
True. That one is a lost cause.
SXX · 5h ago
My company is making pixel art games with high quality art and we released 3 titles so far. While we don't use AI at all we still do experiment with it not to miss what's possible. And so far I just can't agree you can produce anything good consistently with AI.
Pixel art production is not just about making cool lookin pixels, but about consistency. It take a lot of art direction to make professional artists working together to match specific art style.
With 3D games you can mostly use assets from stores and just cover them with effects or shading, but it's just not how pixel art works.
You can generate tons of sprite sheets using AIs and each will look good and sometimes great, but getting right style, lighting and camera perspective is near impossible.
And animations are even harder because there just so little information to work with.
mrob · 6h ago
These are not pixel art because they don't use a regular pixel grid. Additionally, the style is inconsistent, and there are many obvious artifacts. They're also large, which means they don't have to rely so heavily on specialized pixel art techniques. When you're working with images as small as was common in the golden age of pixel art, you have to think carefully about the meaning of every pixel.
Good pixel art is designed around human interpretation of ambiguity. The clearest examine I can think of is in Chrono Trigger. The starting room contains a representation of a typewriter, where the keys are drawn as a checkerboard dither pattern. This doesn't look anything like a real typewriter keyboard. Only the context makes it clear what it is, and in a different context it could be interpreted as a flat expanse of color. I don't think the standard diffusion model architecture is capable of this kind of judgment.
gametorch · 6h ago
I appreciate the honest feedback. Thank you.
Your comment reminds me of this one on Dropbox's Launch HN post:
> I have a few qualms with this app:
> 1. For a Linux user, you can already build such a system yourself quite trivially by getting an FTP account, mounting it locally with curlftpfs, and then using SVN or CVS on the mounted filesystem. From Windows or Mac, this FTP account could be accessed through built-in software.
> 2. It doesn't actually replace a USB drive. Most people I know e-mail files to themselves or host them somewhere online to be able to perform presentations, but they still carry a USB drive in case there are connectivity problems. This does not solve the connectivity issue.
> 3. It does not seem very "viral" or income-generating. I know this is premature at this point, but without charging users for the service, is it reasonable to expect to make money off of this?
So I'll take your comment as a positive signal. I think I'm on to something good :)
Fraterkes · 7h ago
Look, I don’t really like this stuff philosophically, but also in a practical sense I think a lot of these look pretty bad and have obvious ai artifacting.
Why can’t you just find an artist who will work with you on your game? Get someone a job, make a friend, share cross-disciplinary insights and make a better game.
DrSiemer · 6h ago
For most people, especially younger, pixel art is just a retro esthetic. Most indie developers do not even have the money to pay themselves, let alone hire an artist.
Fraterkes · 6h ago
I’m not a pixel purist, I think a lot of these just straightforwardly don’t look great.
The way a lot of small indies get made is a group of people (often friends) with different skillsets work together and share the revenue. “Cofounders” in hn speak.
DrSiemer · 5h ago
We're not that far off from generated pixel art looking indistinguishable from the real deal.
Friends working together generally only happens when they are still in school and have a surplus of time for projects that do not generate a consistent paycheck. As somebody with a lot of ideas and plenty of capable friends, my experience is that none of them are interested in doing unpaid work once they get on the job market.
Fraterkes · 3h ago
I can only judge the current state, which is already being prominently marketed.
When two people start a company, neither is employing the other, and they also aren’t working for free.
strogonoff · 8h ago
What makes art different from an arbitrary pretty thing?
In my view, it is the nature of art as self-expression and metaphorical, transcending the constraints of literal verbal meaning, human-to-human communication. Ergo, when there is no self to express on the other side, it could be a pretty thing, but categorically not art.
gametorch · 8h ago
You know, I agree.
I think there are two components of art that make it valuable - first, the entire history and human context behind the art. This component is sentimental in nature and inextricably linked to both the artist and the beholder. People value art because it means something to them in a way that is often indescribable. And yet you can definitely say that this sometimes-indescribable meaning is definitely not "because it looks cool." It seems related to the sum of all human feeling about the art.
The second component is being automated away. That component is "this configuration of pixels can be used to convey meaning in a practical sense." I need a retro pixelated cactus for my game. I can generate a really good looking one for 10 cents because this component of art has been commodified.
Just because the second component of art's value has been commodified and its value has been driven to zero through competition, that doesn't mean the value of the first component has changed at all. If anything, I'd think people now value art with most of its valued anchored in the first component even more than they used to.
One last thought - I agree with your semantics regarding what we define as art. I apologize for calling this commodified thing, for lack of a better word, "art" throughout my comment. Whatever you call it, I want it for my video games. And I still want the real true human art just as much, if not more, than before the advent of these technologies.
userbinator · 7h ago
first, the entire history and human context behind the art
In some ways, AI art embodies this even more --- it's the sum of innumerable human artists whose work was used in its training, and of course the human that prompted it to generate.
jmkd · 8h ago
The second component you refer to is the craft, or making of artworks. This was automated away over a hundred years ago by Duchamp with 'Fountain' an industrially made urinal.
Since then the making of a work and the meaning of it have become two mostly independent concepts, with neither particularly reliant on the other.
bawolff · 7h ago
That might be nice philosophically, but i don't think that is how the average human defines art.
Even people who like art don't really define it that way afaict. For example, "death of the author" is a hugely popular concept when it comes to art, where the idea is that what matters is what the art makes you feel, not what the author was trying to communicate.
xyzal · 7h ago
Exactly. Artistic and aesthetic values are distinct concepts.
That Mona Lisa replica in your living room has the same aesthetic value than the original, as it is a perfect copy, but vastly inferior artistic value.
Art is defined by context.
Dkumpikolo · 8h ago
I think you find art as you describe it in galleries not in pixel art collections.
And do we all actually care that much? No otherwise it would not be a breadless job.
We don't value artists they struggle financially and most do contract work and not their art.
bawolff · 7h ago
> We don't value artists they struggle financially and most do contract work and not their art.
I think taylor swift is doing fine.
Its not that we don't financially value art as a society, its just that we do it very unequally.
Dkumpikolo · 3h ago
You know exactly that I did not mean the 0.01% of artists.
We do it aligned as a mass phenomenon.
But still Taylor for sure makes good music but how many people really pay for art?
And then also art from a specific artist.
spencerflem · 8h ago
Yeah it makes me sad. I play games to see something neat someone else found and made and wanted to show me. Ai images are just content for its own sake and we have so much of that already.
alisonatwork · 7h ago
I've played a bunch of great RPG Maker and Ren'Py games that used creative commons art and music. The draw of those games was experiencing the story that the creator wanted to tell. What's the difference if the same creator used AI generated art instead? Seems to me it could help creators to express themselves better, not worse.
pjerem · 6h ago
I agree.
But in your case the art would be in everything else that isn’t the sprites. By everything else I mean everything the author would have care about. It wouldn’t mean that the sprites couldn’t be aesthetically appealing but only that this part of the game wouldn’t be what’s artistic about it.
I think IA can be incorporated into art as long as it is made consciously, but IA generated content in itself couldn’t be the art piece.
LoveMortuus · 8h ago
These are really nice! I've been waiting for the AIs to get good enough so that I could use them to generate game art, especially because I can't draw at all and because game dev is just a hobby, I can't spend money on it to hire someone to do the art for me.
keyle · 7h ago
I'm still not sure where I stand with this. It reminds me a bit of the skeuomorphic days, where we went to flat design overnight and thanked the designers goodbye. Developers got tired of waiting for a button sprite to be designed for the website. In a way it was cleaner, leaner and faster. But we lost a lot of good work in the process (I was one of those designers).
I think where I stand is AI for supplementing, yes, for replacing, no. Extend from existing work, yes. But if you distance yourself so far from the art that you don't know how it's made, you'll be flying while running out air.
levmiseri · 6h ago
> It reminds me a bit of the skeuomorphic days, where we went to flat design overnight and thanked the designers goodbye
This didn't happen. Even if you consider only the 'aesthetic' part of design and ignore function (impossible, but for the sake of the argument) — the striving for reduction and removal of unnecessary real-world metaphors was a natural evolution driven by designers, not PMs, not engineers, nor any higher-ups. 'Flat' design was in hindsight inevitable, but you can't just get rid of decorations, ornaments and gradients and call it a day. All major digital products were and are designed by designers. Not by engineers who got tired of waiting for a button sprite.
Cyberdogs7 · 10h ago
Pretty good stuff! I took the opposite journey and started in games then went to big tech, though I have continued to stay with the game industry the whole time with an indie game company I run on the side.
I did take up the challenge of trying to prompt an image generator into giving me a useable 2d sprite <https://nlevel.ai/images/K4oeERN4a0By/view)> and it's much harder than it looks.
I assume you are running some type of LLM to specially format the prompts to the image models, or is it more complex than that?
gametorch · 10h ago
Hey thanks for the feedback! Your image link is giving me 404, fyi.
> I assume you are running some type of LLM to specially format the prompts to the image models, or is it more complex than that?
Yes, we pre-prompt by default, but you can manually disable this. If you disable it you also avoid paying the LLM input/output token tax.
Strangely enough, though, specifically formatting prompts and creating pre-prompts hasn't really had as large of an effect as I thought it would on the quality of the end result.
Despite how little of an effect pre-prompting has, I still think most of the value add right now comes from the business logic. It's about what you do before and after sending the prompt to get the best results. It'll also be about crafting the most frictionless UI for a human. For example, they need to be able to easily convey to the computer "this image and this other image are good in these specific ways, please combine them to make this third image."
The other long term business view here is these models are truly improving in quality exponentially and decreasing in cost exponentially. Read: they are commodities. This means you WILL get categorically replaced if you are in the business of having the best model (the most sentimentally valuable piece of gold! or corn bushel!). The real value add here, if there is any long term, has to be in the UI, organization, and distribution of the commodity. That's what I'm trying to build with GameTorch.
Please leave creative work for human beings and teach "AI" how to do dishes or wash clothes, and other chores, that would leave more time for human beings to do creative work.
rambambram · 5h ago
I just want to play this game in my browser!
seattle_spring · 8h ago
Cool, but these are all static images. Is there a way to turn these into animations, including stuff like rotation (player looking forward, player turned around, etc.)
gametorch · 7h ago
Yes, but it's not perfect yet.
Here are two examples, from the live website. [1][2]
What you described is a really hard problem. I'm working on it every single day.
The bad news is we aren't really there yet. The good news is I have a lot of ideas to try out in my backlog.
Looks nice enough but the pixels are not all the same size which is kinda upsetting.
CommanderData · 6h ago
Call me when AI can do animated full pixel sunsets with moving clouds, birds a stream and make it look magical.
I still enjoy Reddits pixel art sub, and AI generated content doesn't come close.
gametorch · 11h ago
TL;DR:
I started making a video game. I assumed 2D video game art generation was a solved problem. It wasn't. I tried to solve it and ended up getting close. This post's URL provides an example of how far image generation models have come (and how far coding LLMs have come, because they built that entire website, with my assistance, in 2 months).
Long Post:
I originally started programming because, like many others, I wanted to make video games. A bachelor's degree in Computer Science, a few internships in Silicon Valley, and some year-long stints at various companies that shall not be named here -later, I ended up where I started from: making video games.
I set out on my game-making journey intent to fully embrace change, because the times they are a-changing, as they always seem to do. First, I'd write the game in Rust, a language which I wish we had when I was in college (though I admit learning C is absolutely a pedagogical rite in any up-to-snuff CS program). Second, I'd use LLMs to guide me on my way.
I quickly found my bearings with the Bevy game engine [1]. I implemented Flappy Bird for practice. Sailing was smooth. It was time to make the game I originally intended to make.
Early in the development of the game I found the need to procure assets. I had already stumbled upon Meshy [2], a web app that let's you create 3d meshes from prompts. I think this is a great product and I gladly paid for it. However, I was making a 2D game, and 2D games need 2D assets. I assumed the 2D-equivalent of Meshy existed and I could pay for it.
How wrong I was. Alas, I could simply ask ChatGPT to generate 2D sprites for me, no? No. Well, yes and no. Yes, in the sense that ChatGPT does generate some pretty damn good looking sprites. No, in the sense that hours of painful manually cropping and Photoshop work was about to beset me. So I took matters into my own hands and tried to build it myself.
these are very nice, although I wonder if you can improve the consistency image-to-image somehow? The leading cactus image and tree images look like they come from different collections because one is black-outlined and one is not. The kelp pirates are lit differently from each other.
okkdev · 8h ago
Most soulless looking pixel art I've seen in a while...
I don't know what would be a practical way to do it, but I imagine some postprocessing step where a consistent matrix is enforced.
Not adhering to a pixel grid is just… not pixel art.
https://en.wikipedia.org/wiki/Sonic_the_Hedgehog_(1991_video...
Look at the vegetation in the background. However, this is also a good example of the difference between real art and AI slop. There's human intention here, and I can guess a plausible scenario of how it happened.
I suspect the checkerboard pattern in the foreground came first. It has the practical utility of showing off the game's primary selling point, which is Sonic's high speed. You need a high-contrast pattern that's not too distracting and doesn't use too much memory. A checkerboard is an obvious choice. And once you've made that decision, it influences the aesthetics of everything else.
Checkerboards were a very popular texture in early 3D rendering. Any computer artist working in 1990 would be familiar with this, and would at least subconsciously think of 3D graphics. This is likely the reason for the stylized leaves on the trees, which look like low-polygon 3D models. In 1991 it felt futuristic. I believe the blocky look of the background vegetation is intended to convey this same "computer graphics" feeling. Drawing polygons is impossible, but drawing attention to the pixels gives the same impression of something futuristic.
This is the kind of non-obvious artistic decision that I don't believe current AI is capable of. A human intentionally reduced the graphics quality, and that actually made it look more futuristic. It's the kind of thing you see all the time in good art. Close examination reveals detail that deepens your appreciation for it. AI slop only imitates the surface polish of good art without including this deeper meaning. The closer you examine it the worse it looks.
Don't get me wrong, some of these images do look good, they're just not enough for a game.
You can create animated spritesheets with the click of a button. [1]
You can create as many assets as you want to make your game. It's all prompt-driven.
If you have ideas for things you want generated, let me know and I'll run your prompt for you!
1. https://gametorch.app/public/show_him_winking.mp4
What are you doing to make sure your product meets the quality standards of the industry?
It's only a matter of time until the model is sufficient for everyone's standards.
The most puzzling thing to me is that in a forum full of technologists, a significant minority are betting against the advancement of technology.
If you could let me know why you seem so bent on my failure, I would be forever grateful.
https://gametorch.app/commons/131
There's no way to fix this without redrawing it from scratch.
I was too lazy to even type out a full prompt, just clicked the "touch-up" button which feeds in this prompt:
"please upscale and touch up, but keep the transparent background because this is a video game sprite"
And I got this result:
https://gametorch.app/commons/edits/24
That's with today's models. Six months from now they will be exponentially better and cheaper.
Pixel art production is not just about making cool lookin pixels, but about consistency. It take a lot of art direction to make professional artists working together to match specific art style.
With 3D games you can mostly use assets from stores and just cover them with effects or shading, but it's just not how pixel art works.
You can generate tons of sprite sheets using AIs and each will look good and sometimes great, but getting right style, lighting and camera perspective is near impossible.
And animations are even harder because there just so little information to work with.
Good pixel art is designed around human interpretation of ambiguity. The clearest examine I can think of is in Chrono Trigger. The starting room contains a representation of a typewriter, where the keys are drawn as a checkerboard dither pattern. This doesn't look anything like a real typewriter keyboard. Only the context makes it clear what it is, and in a different context it could be interpreted as a flat expanse of color. I don't think the standard diffusion model architecture is capable of this kind of judgment.
Your comment reminds me of this one on Dropbox's Launch HN post:
> I have a few qualms with this app:
> 1. For a Linux user, you can already build such a system yourself quite trivially by getting an FTP account, mounting it locally with curlftpfs, and then using SVN or CVS on the mounted filesystem. From Windows or Mac, this FTP account could be accessed through built-in software.
> 2. It doesn't actually replace a USB drive. Most people I know e-mail files to themselves or host them somewhere online to be able to perform presentations, but they still carry a USB drive in case there are connectivity problems. This does not solve the connectivity issue.
> 3. It does not seem very "viral" or income-generating. I know this is premature at this point, but without charging users for the service, is it reasonable to expect to make money off of this?
So I'll take your comment as a positive signal. I think I'm on to something good :)
Why can’t you just find an artist who will work with you on your game? Get someone a job, make a friend, share cross-disciplinary insights and make a better game.
The way a lot of small indies get made is a group of people (often friends) with different skillsets work together and share the revenue. “Cofounders” in hn speak.
Friends working together generally only happens when they are still in school and have a surplus of time for projects that do not generate a consistent paycheck. As somebody with a lot of ideas and plenty of capable friends, my experience is that none of them are interested in doing unpaid work once they get on the job market.
When two people start a company, neither is employing the other, and they also aren’t working for free.
In my view, it is the nature of art as self-expression and metaphorical, transcending the constraints of literal verbal meaning, human-to-human communication. Ergo, when there is no self to express on the other side, it could be a pretty thing, but categorically not art.
I think there are two components of art that make it valuable - first, the entire history and human context behind the art. This component is sentimental in nature and inextricably linked to both the artist and the beholder. People value art because it means something to them in a way that is often indescribable. And yet you can definitely say that this sometimes-indescribable meaning is definitely not "because it looks cool." It seems related to the sum of all human feeling about the art.
The second component is being automated away. That component is "this configuration of pixels can be used to convey meaning in a practical sense." I need a retro pixelated cactus for my game. I can generate a really good looking one for 10 cents because this component of art has been commodified.
Just because the second component of art's value has been commodified and its value has been driven to zero through competition, that doesn't mean the value of the first component has changed at all. If anything, I'd think people now value art with most of its valued anchored in the first component even more than they used to.
One last thought - I agree with your semantics regarding what we define as art. I apologize for calling this commodified thing, for lack of a better word, "art" throughout my comment. Whatever you call it, I want it for my video games. And I still want the real true human art just as much, if not more, than before the advent of these technologies.
In some ways, AI art embodies this even more --- it's the sum of innumerable human artists whose work was used in its training, and of course the human that prompted it to generate.
Even people who like art don't really define it that way afaict. For example, "death of the author" is a hugely popular concept when it comes to art, where the idea is that what matters is what the art makes you feel, not what the author was trying to communicate.
That Mona Lisa replica in your living room has the same aesthetic value than the original, as it is a perfect copy, but vastly inferior artistic value.
Art is defined by context.
And do we all actually care that much? No otherwise it would not be a breadless job.
We don't value artists they struggle financially and most do contract work and not their art.
I think taylor swift is doing fine.
Its not that we don't financially value art as a society, its just that we do it very unequally.
We do it aligned as a mass phenomenon.
But still Taylor for sure makes good music but how many people really pay for art?
And then also art from a specific artist.
But in your case the art would be in everything else that isn’t the sprites. By everything else I mean everything the author would have care about. It wouldn’t mean that the sprites couldn’t be aesthetically appealing but only that this part of the game wouldn’t be what’s artistic about it.
I think IA can be incorporated into art as long as it is made consciously, but IA generated content in itself couldn’t be the art piece.
I think where I stand is AI for supplementing, yes, for replacing, no. Extend from existing work, yes. But if you distance yourself so far from the art that you don't know how it's made, you'll be flying while running out air.
This didn't happen. Even if you consider only the 'aesthetic' part of design and ignore function (impossible, but for the sake of the argument) — the striving for reduction and removal of unnecessary real-world metaphors was a natural evolution driven by designers, not PMs, not engineers, nor any higher-ups. 'Flat' design was in hindsight inevitable, but you can't just get rid of decorations, ornaments and gradients and call it a day. All major digital products were and are designed by designers. Not by engineers who got tired of waiting for a button sprite.
I did take up the challenge of trying to prompt an image generator into giving me a useable 2d sprite <https://nlevel.ai/images/K4oeERN4a0By/view)> and it's much harder than it looks.
I assume you are running some type of LLM to specially format the prompts to the image models, or is it more complex than that?
> I assume you are running some type of LLM to specially format the prompts to the image models, or is it more complex than that?
Yes, we pre-prompt by default, but you can manually disable this. If you disable it you also avoid paying the LLM input/output token tax.
Strangely enough, though, specifically formatting prompts and creating pre-prompts hasn't really had as large of an effect as I thought it would on the quality of the end result.
Despite how little of an effect pre-prompting has, I still think most of the value add right now comes from the business logic. It's about what you do before and after sending the prompt to get the best results. It'll also be about crafting the most frictionless UI for a human. For example, they need to be able to easily convey to the computer "this image and this other image are good in these specific ways, please combine them to make this third image."
The other long term business view here is these models are truly improving in quality exponentially and decreasing in cost exponentially. Read: they are commodities. This means you WILL get categorically replaced if you are in the business of having the best model (the most sentimentally valuable piece of gold! or corn bushel!). The real value add here, if there is any long term, has to be in the UI, organization, and distribution of the commodity. That's what I'm trying to build with GameTorch.
Here are two examples, from the live website. [1][2]
What you described is a really hard problem. I'm working on it every single day.
The bad news is we aren't really there yet. The good news is I have a lot of ideas to try out in my backlog.
1. https://gametorch.app/public/show_him_winking.mp4
2. https://gametorch.app/commons/spritesheets/1
Looks nice enough but the pixels are not all the same size which is kinda upsetting.
I still enjoy Reddits pixel art sub, and AI generated content doesn't come close.
I started making a video game. I assumed 2D video game art generation was a solved problem. It wasn't. I tried to solve it and ended up getting close. This post's URL provides an example of how far image generation models have come (and how far coding LLMs have come, because they built that entire website, with my assistance, in 2 months).
Long Post:
I originally started programming because, like many others, I wanted to make video games. A bachelor's degree in Computer Science, a few internships in Silicon Valley, and some year-long stints at various companies that shall not be named here -later, I ended up where I started from: making video games.
I set out on my game-making journey intent to fully embrace change, because the times they are a-changing, as they always seem to do. First, I'd write the game in Rust, a language which I wish we had when I was in college (though I admit learning C is absolutely a pedagogical rite in any up-to-snuff CS program). Second, I'd use LLMs to guide me on my way.
I quickly found my bearings with the Bevy game engine [1]. I implemented Flappy Bird for practice. Sailing was smooth. It was time to make the game I originally intended to make.
Early in the development of the game I found the need to procure assets. I had already stumbled upon Meshy [2], a web app that let's you create 3d meshes from prompts. I think this is a great product and I gladly paid for it. However, I was making a 2D game, and 2D games need 2D assets. I assumed the 2D-equivalent of Meshy existed and I could pay for it.
How wrong I was. Alas, I could simply ask ChatGPT to generate 2D sprites for me, no? No. Well, yes and no. Yes, in the sense that ChatGPT does generate some pretty damn good looking sprites. No, in the sense that hours of painful manually cropping and Photoshop work was about to beset me. So I took matters into my own hands and tried to build it myself.
1. https://bevy.org/
2. https://www.meshy.ai/discover