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.
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.
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.
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?
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.
Geniune question: did you ever tried making pixeled games? Or maybe stack existing pictures into a bigger enviroment picture?
The main issue I can see is the pixel inconsistency. Since sprites move around the screen, whatever misaligments could become noticible in between sprites would be noticible.
The most noticible thing that would make all this sprites unusable in one game — cacti don't fit. Like, visually, they pop. Useful if you make a game about collecting cacti, maybe, since "popping" something will make it appear important.
Another huge problem — all your man have wildely different proportions. It looks jarring, and you will have to rig them into game engine separetly. Spaking of, you will have problems describing what a collision with an object is to an engine the second you will try to add the sprites to a game. People usually just use the non-transparent part, but you have floating almost-transparent pixels in the mix
So if you make this thing for game devs, it would be a great idea to try and make at least one basic game with them. You can always work out kinks that way.
Also, try to make styles more consistent. Most pixel devs already have a style in mind (and most have some sprites as well), and you had to generate in that style perfectly from first few tries, or else they will just draw it.
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.
#/media/File:MD_Sonic_the_Hedgehog.png){kind=link}
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.