Here's a 7388 x 5097 photo I took from a friend's Cessna of the Mission from a similar angle in 2018 which provides a nice point of reference for comparing the fine details:
You can see some buildings, like 350 Treat Ave, which is a white warehouse in the middle of the photo, are blended from two colors in this isometric viewer. It's yellow on the left and grey on the right, when in fact it is uniformly white throughout. This is probably due to image generation running separately on two tiles and then blending them together.
There's almost a hundred of them and I'm not exagerating:
> https://isometriccities.com/ Since Isometric NYC, makers around the world have been rebuilding their hometowns as giant SimCity-style maps. We collect every one of them. 98 maps
I mention this a lot: the two basic post-processing steps you have to do, at a minimum, to give genai images even a semblance of pixel art are palette reduction and square grid alignment.
Unfortunately, the definition of "pixel art" like "vibe coding" has been diluted to the point of almost being meaningless.
If you have a project like this with tens of thousands of tiles, you have to run them all through square grid alignment. That process tries to pick an optimal square grid and then reduce the image to a discrete set of pixels, with each square resolved to a single color.
It's similar to what you do when you shrink an image using nearest-neighbor scaling. As a result, you mitigate a lot of the irritating high-frequency noise and fringing that occur in these higher-resolution pixel art images from generative AIs.
Applying it to a project with tens of thousands of images still requires a lot of manual intervention. This is what makes separates good pixel art from bad pixel art: even at a low resolution of 16x16 or 32x32, it's still recognizable.
Try taking any of the 512x512 tiles from this particular isometric city project (or any of the other dozen variants) and feed'em into Spritefusion and you'll quickly see what I mean.
Why does anyone need to publish anything like this snapper, the person doing pixelart, if he understands pixelart, should know to do it himself! this is the third map published here in apparent diffusionart which is absolutely not a pixelart. besides for buildings - pixelart artists would choose repeating stuff that diffusion models would never be able to reproduce, at least for the FP errors that pile while it happens.
this is really cool, but I wish it was oriented north south (maybe thats not as pleasing re the isometry?), just because that's how I think about the city and navigating it, so this makes me like, instinctually rotate my head.
How much it cost you paint per km2 after training the model? I did (and publish here) https://www.tinyvalencia.com and got to 0.4$ per km2 after training the model (for 15$ or so).
I've been seeing a lot of these kinds of projects make the rounds on Twitter lately. Yours is just as good as the best of them, so great job making it more efficient without loss of quality!
But... I'm always disappointed when I zoom in and see that it's not actually pixel art, and it's full of the kinds of artifacts that are typical of any AI-generated imagery :(
Thank you! I tried to get Fable to take a stab at writing the technicals in the original reflection post, but I really didn't like it and sought to rewrite it myself. Took a week longer, but I'm way happier with the result.
Not bad. I feel like instead of using an image-mode model, it would be better to use a code model that generates the scene. Reason I think this is that the image generators aren't respecting the rules of "pixel art". Straight lines and flat planes are straight and flat in pixel art. These images are wavy and irregular.
This is neat but maybe we could call it "pixel-art-styled?"
OP didn't mention it (or maybe I missed it), but the underlying imagery comes from the Google 3D photorealistic tiles API: https://developers.google.com/maps/documentation/tile/3d-til...
They offer a generous 1000 API calls per month for free, in case anyone is interested in replicating this.
Here's a 7388 x 5097 photo I took from a friend's Cessna of the Mission from a similar angle in 2018 which provides a nice point of reference for comparing the fine details:
https://i.dllu.net/2018-04-14-23-31-28_DSC04018_6201e8e3a215...
You can see some buildings, like 350 Treat Ave, which is a white warehouse in the middle of the photo, are blended from two colors in this isometric viewer. It's yellow on the left and grey on the right, when in fact it is uniformly white throughout. This is probably due to image generation running separately on two tiles and then blending them together.
Still, from a distance, this looks great.
Sorry to be a downer, but this looks really bad to me. It completely lacks the sumptuous precision of real pixel art.
I was going to post the same, zooming in it's jpg artifacts and ghosting, I assume it was pixel perfect at some point in the process but isn't now.
…not actually pixel art though, no?
Search for "pixel art" in HN's search and you'll see most of the last submissions are the same fake pixel art.
"pixel art" NYC: https://news.ycombinator.com/item?id=46721802
"pixel art" Rio: https://news.ycombinator.com/item?id=50448289
"pixel art" London: https://news.ycombinator.com/item?id=50163735
"pixel art" Valencia: https://news.ycombinator.com/item?id=50441518
"pixel art" world: https://news.ycombinator.com/item?id=49674973
There's almost a hundred of them and I'm not exagerating:
> https://isometriccities.com/ Since Isometric NYC, makers around the world have been rebuilding their hometowns as giant SimCity-style maps. We collect every one of them. 98 maps
Yeah. I'm surprised this technique hasn't yet caught on among people doing generative visuals with "pixel art" aesthetics: https://github.com/Hugo-Dz/spritefusion-pixel-snapper
I mention this a lot: the two basic post-processing steps you have to do, at a minimum, to give genai images even a semblance of pixel art are palette reduction and square grid alignment.
Unfortunately, the definition of "pixel art" like "vibe coding" has been diluted to the point of almost being meaningless.
If you have a project like this with tens of thousands of tiles, you have to run them all through square grid alignment. That process tries to pick an optimal square grid and then reduce the image to a discrete set of pixels, with each square resolved to a single color.
It's similar to what you do when you shrink an image using nearest-neighbor scaling. As a result, you mitigate a lot of the irritating high-frequency noise and fringing that occur in these higher-resolution pixel art images from generative AIs.
Applying it to a project with tens of thousands of images still requires a lot of manual intervention. This is what makes separates good pixel art from bad pixel art: even at a low resolution of 16x16 or 32x32, it's still recognizable.
Try taking any of the 512x512 tiles from this particular isometric city project (or any of the other dozen variants) and feed'em into Spritefusion and you'll quickly see what I mean.
https://imgpb.com/ePXki
Mixels and incoherent palettes had confused the term pixel art long before genAI imo.
Why does anyone need to publish anything like this snapper, the person doing pixelart, if he understands pixelart, should know to do it himself! this is the third map published here in apparent diffusionart which is absolutely not a pixelart. besides for buildings - pixelart artists would choose repeating stuff that diffusion models would never be able to reproduce, at least for the FP errors that pile while it happens.
This is so beautiful! Something soothing and ultra clean about it. Maybe all the maps should have this rendered version as default.
this is really cool, but I wish it was oriented north south (maybe thats not as pleasing re the isometry?), just because that's how I think about the city and navigating it, so this makes me like, instinctually rotate my head.
Thanks! I choose this orientation to highlight the coast of the SF Bay area more.
That's cool! The coast really frames it nicely.
Also see isometric Toronto: https://iso.sloptoronto.ca/#v=3&x=102672&y=64255&z=0.262
Wow, it captured the pool in my apartment complex. I love this! I was a bit disoriented by north being left.
How much it cost you paint per km2 after training the model? I did (and publish here) https://www.tinyvalencia.com and got to 0.4$ per km2 after training the model (for 15$ or so).
It cost around $0.09 per km2 after training the model
I've been seeing a lot of these kinds of projects make the rounds on Twitter lately. Yours is just as good as the best of them, so great job making it more efficient without loss of quality!
But... I'm always disappointed when I zoom in and see that it's not actually pixel art, and it's full of the kinds of artifacts that are typical of any AI-generated imagery :(
Awesome. Next step: Make the trains and cars move!
Found my house in East Bay and it looks accurate, cool.
This is my favorite one yet! I used to live in SoMa and this captured my apartment and the nearby Caltrain wonderfully!
Is there any way to rotate the perspective, as in Google Maps with 2-finger gestures (touchscreens) or shift-drags (mice/etc)?
(I couldn't find it, if it already exists.)
Not at the moment. It's captured at a fixed angle.
Hi Justin, great writeup!
I really appreciate your human-written documentation because it's nice to read your thought process instead of LLM marketing soundbites.
re: orientation. I know it's creator's choice, but for me, it's either north or west towards the ocean. East makes me want to hold shift to rotate. :)
Thanks for making and sharing this!
Thank you! I tried to get Fable to take a stab at writing the technicals in the original reflection post, but I really didn't like it and sought to rewrite it myself. Took a week longer, but I'm way happier with the result.
Technicals written by Fable at the end of this post: https://www.justinchavez.com/blog/isometric-sf/
This style seems achievable while retaining 3D rendering if the tiles are preprocessed in 3D instead of 2D and some clever shaders are used.
unfortunate that google explicitly forbids 3D model generation from their map data
I'm pretty sure that re-rendering the tiles to 2D and rehosting them on your own website is also forbidden. So no difference there.
Just wanted to say... this is beautiful!
This is amazing and is 1:1 global map. Did you use google maps API to build this or UE5?
I used the Google 3D Photorealistic Tiles API then rendered the tiles in three.js to capture them at the angle I wanted
Google 3D tiles.
Not bad. I feel like instead of using an image-mode model, it would be better to use a code model that generates the scene. Reason I think this is that the image generators aren't respecting the rules of "pixel art". Straight lines and flat planes are straight and flat in pixel art. These images are wavy and irregular.
Impressive! And the aesthetics look great!