I don't think your algorithm is correct. At least on the checkerboard example on the cube face the diagonals are curved. Perspective transformation doesn't do that.
Possibly you do the subdivisions along the edges uniformly in the target space, and map them to uniform subdivisions in the source space, but that's not correct.
edit:
Comparison of the article's and the correct perspective transform:
Completely backwards. Math is much more difficult than programming and LLMs still can't consistently add numbers correctly last I checked. What a strange attitude to take.
Even more obviously, the squares in the front aren’t bigger than the squares in the back. It looks like each square has equal area even as their shapes change.
It’s fascinating how plausible it looks at a glance while being so glaringly wrong once you look at it more closely.
Author here: I don’t think the commenter here has set the same focal length, the focal length can make a surface appear curved, I set it explicitly to a low value to test the algorithm’s ability to handle the increased distortion. You can google “focal length distortion cube” to see examples of how a focal length distorts a grid or you can google “fish eye lens cube” etc.
Edit: I think there’s a lot of confusion because the edges of the cube (the black lines), do not incorporate the perspective transform all along their edge. The texture is likely correct given the focal length, and the cube’s edge is misleadingly straight. My bad, the technique is valid, but the black lines of the cube’s edge are misleadingly straight (they are not rendered the same way as the texture)
I think the original commenter is correct that there is a mistake in the perspective code.
It seems the code calculates the linear interpolation for the grid points too late. It should be before projecting, not after.
I opened an issue ticket on the repository with a simple suggested fix and a comparison image.
That admittedly looks a lot more correct! Thanks for digging in, i will absolutely test and submit a correction to the article (i am still concerned the straight edges are misleading here)! And thanks to the original commentor as well!
I think I will try to quickly output an animated version of each subdivision level, the animation would make it a lot more clear for me!
I might be missing something but you sound genuinely confused to me. The perspective in your post is linear perspective. It's the one used in CSS and it doesn't curve straight lines/planes. It's not the perspective of fish-eye images (curvilinear perspective).
I was at least a little confused because yea fish eye isn’t possible with a 4x4 perspective transform matrix. I’m investigating an issue with the projection thanks to some help from commenters and there will be a correction in the article, as well as an animation which should help confirm the projection code.
It is possible to approximate perspective using piecewise affine transformations. It is certainly possible to match the perspective transformation at the vertices of the subdivisions, and only be somewhat off within.
With 6 degrees of freedom, you can only fit 3 2d points at a time. Triangulation causes the errors shown in the article, hence why subdivision is needed.
I think GP's point is that besides the unavoidable distortions coming from approximating a perspective transform by a piece-wise affine transform, the implementation remains incorrect.
It’s worth noting that this same restriction of not being able to do perspective transformations is also one of the defining characteristics of PlayStation 1 graphics. And the workaround of subdivision is also the same workaround PS1 games used.
It is also a limitation that many initial DOS 3D software rasterized games had (e.g. Descent.)
This is because perspective transform requires a divide per pixel and it was too costly on the CPUs of the time, so they skipped it to get acceptable performance.
Huh that’s so crazy. I had that in my head as I was reading the article. I was thinking about some car game and the way the panels would look when it rotated in your “garage”.
A friend was writing a flight simulator from scratch (using Foley and van Dam as reference for all the math involved). A classic perspective problem might be a runway.
Imagine a regularly spaced dashed line down the runway. If you get your 3D renderer to the stage that you can texture quads with a bitmap, it might seem like a simple thing to have a large rectangle for the runway, a bitmap with a dashed line down the center for the texture.
But the texture mapping will not be perspective (well, not without a lot of complicated math involved).
Foley and van Dam say — break the runway into a dozen or so "short" runways laid end to end (subdivide). The bitmap texture for each is just a single short stripe. Now because you have a bunch of these quads end to end, it is as if there is a longer runway and a series of dashed lines. And while each individual piece of the runway (with a single stripe), is not in itself truly perspective, each quad as it gets farther from you is nonetheless accounting for perspective — is smaller, more foreshortened.
Back in the early 90s I did a version of Bresenham's algorithm that would rasterize the hyperbolic curves that perspective-correct texture mapping required. It worked correctly though the technique of just doing a division every n pixels and linearly interpolating won out in the end, if I recall.
You could also avoid divisions entirely, while still keeping 100% correct perspective, by "rasterizing" the polygon following the line of constant Z.
You would save the divs, but then you would draw mostly outside the cache, so not a panacea, but for large surfaces it was noticeably nicer than divide-every-N-pixcels approximation.
The goal for this vanilla TS renderer is to have visual diffing on GitHub and a renderer that works without a browser environment. Most 3D renderers focus on realtime speed, not file size and runtime portability. I think in practice we will configure the subdivisions at something like 64 for a good file size tradeoff
This doesn't answer the question. If you're doing all this work in JS to render a static SVG, why not just "do it right" and output a static PNG instead?
The top of the PCB (the lines etc) are computed as an SVG, i would have to have an SVG rasterizer just to begin with that approach, then would be limited by what images I could rasterize. It would also be much much slower than quickly computing matrices
I was going to suggest raylib for server-side rendering, but it adds a non-JS dependency. Apparently it has optional support for rendering SVGs to textures.
I was on the original SVG team at Adobe back in '00 and built some of the first public demos that used the technology. This kind of 3d work was some of the first stuff I tried to do and found it similarly lacking due to the lack of possible transforms. I had some workarounds of my own.
One demo had a 3d stack of floors in a building for a map. It used an isometric perspective (one where parallel lines never converge) and worked pretty well. That is pretty easy and can be accomplished with rotation and scaling transforms.
The other was a 3d molecule viewer where you could click and drag around to view the structure. This one basically used SVG as a canvas with x and y coordinates for drawing. All of the 3d movement was done in Javascript, computing x and y coordinates and updating shapes in the SVG DOM. Styles were used to handle single / double / triple bonds, and separate groups were used to layer everything for legibility.
I hope someday where we get back to a simple HTML/CSS standard for "text" pages and that's it. No JavaScript, no DOM. This covers 70% of the web use cases.
"Everything else" would be a pluggable execution runtime that are distributed as browser plugins: [WASM Engine, JVM engine, SPIR-V Engine, BEAM Engine, etc] with SVG as the only display tech. The last thing we'd define is an interrupt and event model for system and user interactions.
What does he think SVG is doing under the hood? Rasterization. Everything does rasterization at some point in the process. Calculating 512 clip paths to render a single quad that could be drawn in a single for loop is insane.
Your renderer looks awesome! I was surprised there wasn't an "off the shelf" SVG renderer in native TS/JS, it's a big deal to be able to create 3D models without a heavy engine for visual snapshot testing!
The implementation shown in the video is actually particularly slow because all the geometric transformations are implemented in terms of lenses/optics ([1]) and ramdajs ([2]). So the whole mess is a gigantic stack of nested, composed and curried functions, instead of raw linear algebra (just becaus I could).
I later optimized the hotpath and it is significantly faster (still miles behind webgl/webgpu obviously). You can try yourself if you scroll alll the way to the veeeerrrry bottom here [3].
An other approach would be to apply the transformation to SVG elements separately. Inkscape has a perspective transformation tool, which you can apply to paths (and paths only). It probably needs to do approximation and subdivision on the path itself though, which is possibly more complex.
One possibly uncalled-for piece of feedback: is that USB-C connection finished, and is it complying with the various detection resistor requirements for the CCx pins? It seemed very bare and empty, I was expecting some Rd network to make the upstream host able to identify the device. Sorry if I'm missing the obvious, I'm not an electronics engineer.
Because it’s only being used for power and doesn’t need a lot of power, it works for the simple board we rendered. In practice you would absolutely want to set the CC1 and CC2 configuration with resistors!
This is a cool project and I think I can use that. I was just wondering if perspective correctness was all that important for a PCB renderer? The distortion should be minimal for these kind of images and I think old CAD programs often did not use correct perspective as well.
Defs is also how the arrows work in this WebGL2 diagram[2], and in fact, I don't think they're possible without defs, because of `marker-end` which seems to require a marker present in defs.
Defs saved the day here on file size- repeating the image (which we usually base64 encode) would have caused a much larger file size and made rasterization much more appealing!
> Very small files and a much simpler rendering scheme!
For a 400x400 SVG with 6 surfaces and 64 subdivisions your file size is only 10x smaller than an uncompressed bitmap. Your SVG should scale linearly with number of objects and be constant with resolution, while an image would scale with the resolution (quite favorably if compressed) and be constant with the number of objects. I'd be interested to know the size of the example at the top of the article.
Also you already have the math to transform points!
> I don’t have to rasterize my SVGs the represent the top of my board.
Interesting, I've been doing 3D SVG by storing the xyz-coordinates in a separate array and using inlined javascript to calculate & refresh the 2D coordinates of the SVG items themselves after rotation. But this means that the file only works in a browser. Maybe it could be possible to replace the javascript with native functions, so the same file would work everywhere.
Yeah, paths are saved in an array where each path segment is a list of control points coupled with the corresponding path command (M, L, C). Those can be used to recreate the path item.
This is nice, but the article left me unconvinced that you need textures at all.
Be it a checker or the drawing on a circuit board, can't you keep everything as vectors, thus avoiding the problem entirely?
Circuit boards have holes, cutouts and import STL/OBJ components that we'll eventually support in this 3d renderer. Assuming we get that far I may have to rename it from "simple-3d-svg"!
I think you'll probably run into performance problems with SVG before you get too far. I can't imagine SVG will perform fluidly with complex circuit boards.
SVG elements are DOM elements after all, and too many DOM elements will cause browser performance issues. I know this the hard way, after adding a few hundred SVG <path> elements with a few hundred <div> elements in a React-based interactive web application, I ended up needing to move to a canvas solution instead, which works amazingly well.
I really hope you have all that figured out, because I don't think it's going to work well using SVG to render complex circuit boards. But maybe your product is only working with very simple circuit boards?
I’m not certain, but I think Firefox just implemented 3D transformations for SVG from the start. It wasn’t exactly hard to conceive. Certainly by mid-2017 it had it. Somewhere around that time there was also concerted effort toward aligning SVG and CSS.
(Firefox’s implementation does still suffer from one long-standing bug which means you want to make sure your viewbox unit is larger than one device pixel, but that’s normally not hard to achieve. https://oreillymedia.github.io/Using_SVG/extras/ch11-3d.html... shows what it’s about. I don’t really understand why that problem isn’t fixed yet; what I presume is the underlying issue affects some HTML constructs too when you scale things up, and surely it’s not that rare? I know I found one such problem a decade ago (and, being in HTML, it couldn’t be worked around like you can with SVG). They’ve improved things a bit, but not entirely.)
Sadly, no one else seemed all that interested in making 3D transformations work properly in SVG content.
Three.js has had an SVG rendering back end for 13 years. It's going to be pretty hard to get much more popular than Three.js to get over the browser vendors' reluctance to make any changes to SVG.
Possibly you do the subdivisions along the edges uniformly in the target space, and map them to uniform subdivisions in the source space, but that's not correct.
edit:
Comparison of the article's and the correct perspective transform:
https://imgur.com/RbRuGxD
Completely backwards. Math is much more difficult than programming and LLMs still can't consistently add numbers correctly last I checked. What a strange attitude to take.
It’s fascinating how plausible it looks at a glance while being so glaringly wrong once you look at it more closely.
Edit: I think there’s a lot of confusion because the edges of the cube (the black lines), do not incorporate the perspective transform all along their edge. The texture is likely correct given the focal length, and the cube’s edge is misleadingly straight. My bad, the technique is valid, but the black lines of the cube’s edge are misleadingly straight (they are not rendered the same way as the texture)
I opened an issue ticket on the repository with a simple suggested fix and a comparison image.
https://github.com/tscircuit/simple-3d-svg/issues/14
More reading: https://retrocomputing.stackexchange.com/questions/5019/why-...
This is because perspective transform requires a divide per pixel and it was too costly on the CPUs of the time, so they skipped it to get acceptable performance.
It's funny that, in today's CPUs, floating point divide is so much faster than integer divide.
A friend was writing a flight simulator from scratch (using Foley and van Dam as reference for all the math involved). A classic perspective problem might be a runway.
Imagine a regularly spaced dashed line down the runway. If you get your 3D renderer to the stage that you can texture quads with a bitmap, it might seem like a simple thing to have a large rectangle for the runway, a bitmap with a dashed line down the center for the texture.
But the texture mapping will not be perspective (well, not without a lot of complicated math involved).
Foley and van Dam say — break the runway into a dozen or so "short" runways laid end to end (subdivide). The bitmap texture for each is just a single short stripe. Now because you have a bunch of these quads end to end, it is as if there is a longer runway and a series of dashed lines. And while each individual piece of the runway (with a single stripe), is not in itself truly perspective, each quad as it gets farther from you is nonetheless accounting for perspective — is smaller, more foreshortened.
It was avoided in the Foley and Van Dam days because it requires a division per rasterized pixel, which was very slow in the late 80s.
Meanwhile.. drawing 512 subdivisions for a single textured quad.
It's a cute trick, certainly, but ask this thing to draw anything more than a couple thousand elements and I bet it's going to roll over very quickly.
Just use webgl where perspective-correct texture mapping is built into the hardware.
https://github.com/raysan5/raylib/discussions/3741
I was on the original SVG team at Adobe back in '00 and built some of the first public demos that used the technology. This kind of 3d work was some of the first stuff I tried to do and found it similarly lacking due to the lack of possible transforms. I had some workarounds of my own.
One demo had a 3d stack of floors in a building for a map. It used an isometric perspective (one where parallel lines never converge) and worked pretty well. That is pretty easy and can be accomplished with rotation and scaling transforms.
The other was a 3d molecule viewer where you could click and drag around to view the structure. This one basically used SVG as a canvas with x and y coordinates for drawing. All of the 3d movement was done in Javascript, computing x and y coordinates and updating shapes in the SVG DOM. Styles were used to handle single / double / triple bonds, and separate groups were used to layer everything for legibility.
"Everything else" would be a pluggable execution runtime that are distributed as browser plugins: [WASM Engine, JVM engine, SPIR-V Engine, BEAM Engine, etc] with SVG as the only display tech. The last thing we'd define is an interrupt and event model for system and user interactions.
[1]: https://youtu.be/kCNHQkG1Q24?si=3VxfVFtG2MiEEmlX
I later optimized the hotpath and it is significantly faster (still miles behind webgl/webgpu obviously). You can try yourself if you scroll alll the way to the veeeerrrry bottom here [3].
[1]: https://github.com/calmm-js/partial.lenses [2]: https://ramdajs.com/ [3]: https://static.laszlokorte.de/svatom/
One possibly uncalled-for piece of feedback: is that USB-C connection finished, and is it complying with the various detection resistor requirements for the CCx pins? It seemed very bare and empty, I was expecting some Rd network to make the upstream host able to identify the device. Sorry if I'm missing the obvious, I'm not an electronics engineer.
See [1] for instance.
[1]: https://medium.com/@leung.benson/how-to-design-a-proper-usb-...
[2] https://webgl2fundamentals.org/webgl/lessons/resources/webgl...
For a 400x400 SVG with 6 surfaces and 64 subdivisions your file size is only 10x smaller than an uncompressed bitmap. Your SVG should scale linearly with number of objects and be constant with resolution, while an image would scale with the resolution (quite favorably if compressed) and be constant with the number of objects. I'd be interested to know the size of the example at the top of the article.
Also you already have the math to transform points!
> I don’t have to rasterize my SVGs the represent the top of my board.
Ahhhhhh. This clears it all up!
SVG elements are DOM elements after all, and too many DOM elements will cause browser performance issues. I know this the hard way, after adding a few hundred SVG <path> elements with a few hundred <div> elements in a React-based interactive web application, I ended up needing to move to a canvas solution instead, which works amazingly well.
I really hope you have all that figured out, because I don't think it's going to work well using SVG to render complex circuit boards. But maybe your product is only working with very simple circuit boards?
(Firefox’s implementation does still suffer from one long-standing bug which means you want to make sure your viewbox unit is larger than one device pixel, but that’s normally not hard to achieve. https://oreillymedia.github.io/Using_SVG/extras/ch11-3d.html... shows what it’s about. I don’t really understand why that problem isn’t fixed yet; what I presume is the underlying issue affects some HTML constructs too when you scale things up, and surely it’s not that rare? I know I found one such problem a decade ago (and, being in HTML, it couldn’t be worked around like you can with SVG). They’ve improved things a bit, but not entirely.)
Sadly, no one else seemed all that interested in making 3D transformations work properly in SVG content.
Why did you feel you had to do this with SVG?