Skip to content
/ qqlrs Public

Rust-based QQL renderer with features and speed

License

Apache-2.0, MIT licenses found

Licenses found

Apache-2.0
LICENSE-APACHE
MIT
LICENSE-MIT
Notifications You must be signed in to change notification settings

qql-art/qqlrs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

qqlrs

Alternative renderer for QQL artwork.

This is a from-scratch re-implementation of the QQL algorithm, ported fairly directly from the original JavaScript source to Rust by @wchargin. Some notable characteristics:

  • Compatibility: This implementation is entropy-compatible with the original, using the same RNG algorithms and sampling. A given QQL seed should look the same in both implementations (see "Fidelity expectations" for details).

  • Performance: Between a faster core (sequential) implementation and support for multicore rendering of a single QQL, this renderer often runs 4–8× faster than the original.

  • Features: This implementation has some additional knobs and dials, for speed, quality, flexibility, and fun. See "Features" below.

  • Hackability: This code is not obfuscated or minified. Structures and dataflow are explicit. There's no hidden or global state. Some components have unit tests. Try poking around: read it, play with it, dump intermediate data structures, write custom searchers, override variables, tweak it yourself. (Just don't suggest that outputs from a tweaked algorithm are "official" QQLs, please.)

As a quick, imperfect performance comparison, rendering some QQLs at 9600×12000px on my laptop, across the original JavaScript renderer, this implementation in sequential mode, and this implementation in parallel mode (--chunks 2x4):

Piece Original Sequential Parallel Speedup
QQL #237 31.57s 9.84s 5.86s 5.4× faster
QQL #200 53.12s 22.88s 8.04s 6.6× faster
QQL #198 48.97s 18.26s 6.02s 8.1× faster

In addition to the speed improvements, this implementation tends to use about half as much RAM compared to the original.

Attribution

The QQL project and the original art algorithm were created by Tyler Hobbs and Dandelion Indigo Lee. This implementation was written by @wchargin.

This repository is dual-licensed under Apache 2.0 and MIT terms.

Getting started

This project uses the standard Rust toolchain. You will need Rust 1.67.0 or newer. To build:

$ cargo test --release
$ cargo build --release

This creates a binary target/release/qql-cli, which you can invoke. In its most basic form, simply pass a seed that you want to render:

$ ./target/release/qql-cli 0x33c9371d25ce44a408f8a6473fbad86bf81e1a178c012cd49a85ffff14c54b46

A close-up of QQL #85, showing the face of the creature.

Pass any additional options after the seed. For example, to render a high-resolution version of just the "face" of QQL #85, with 8 threads of parallelism, and save it to a temporary file:

./target/release/qql-cli \
    0x7d265f38b1d92b48997620b050cf0c534e1908fc203f633b2894ffff10e10c55 \
    --viewport 0.242x0.380+0.378+0.521 \
    --width 19200 \
    --min-circle-steps 32 \
    --chunks 2x4 \
    -o /tmp/out.png \
    ;

Despite rendering a 42 megapixel view into a 461 megapixel virtual canvas, this completes in just 1.7 seconds on my laptop, using 399 MB of RAM at peak.

To learn about what options are available, pass -h to read help text, or keep reading this document.

As a shortcut to build and run together, you can use:

$ cargo run --release -- <seed> [<options>...]

(Don't forget the extra two hyphens after --release.)

Features

Jump:

Viewport restriction

TL;DR: Use the --viewport <WxH+X+Y> option to render just part of the canvas, where each input ranges from 0.0 to 1.0.

If you want just a section of a QQL, the obvious idea is to render the full thing and then crop it. But there's a much faster way, which this implementation provides: restrict the viewport while rendering, so that we don't have to paint all the circles just to throw most of them away.

A single messy circle taken from the center of a QQL that is 160 thousand times larger than this viewport. This image rendered in 2 seconds.

With this technique, you can easily blow up small sections of QQLs to huge resolutions, even if rendering the full image at that resolution would be practically impossible. It's easy to zoom into a QQL far enough that the full image would have trillions of pixels, consume terabytes of RAM, and take days or longer to render. Yet a small window into this virtual canvas can be rendered in under a second.

Use the --viewport <WxH+X+Y> option to specify what section of the canvas you want to render, where each component is a number between 0.0 and 1.0 specifying the width, height, x-offset (from left), and y-offset (from top) of the viewport, respectively. For example, --viewport 0.1x0.2+0.3+0.4 specifies a viewport that is 10% as wide and 20% as tall as the original canvas, with its top-left corner corner 30% of the way in from the left and 40% of the way down from the top. Its bottom-right corner is thus 40% of the way in from the right and 60% of the way down from the top.

Note that when using the --viewport option, the --width still corresponds to the virtual canvas width. So, for instance, a QQL with --width 10000 will be 10000px wide and 12500px tall. If you add --viewport 0.1x0.1+0+0 to render just the top-left corner, the resulting image will be 1000px wide and 1250px tall.

A convenient way to visually find your viewport coordinates is to render a small version of your seed, import it into an image editor like GIMP, and rescale it to 1000×1000px (which will distort the aspect ratio; that's okay). Then, use the crop tool to draw a rectangle over the viewport that you want to render, and just read off the position and size from the side menus. Divide by 1000 to get the relative numbers. For example, with QQL #237:

Screenshot of GIMP with a QQL resized to 1000×1000 and an active crop window, whose numeric coordinates are highlighted.

(If you prefer not to distort the aspect ratio, you can resize to 1000×1250px instead. Just divide the vertical coordinates by 1250 instead of 1000.)

Multicore rendering

TL;DR: If your CPU has (say) 8 threads, add --chunks 2x4 to render most QQLs faster.

Any headless renderer makes it easy to render many QQLs in parallel: simply start multiple copies of the renderer, tasked with different seeds. But this implementation also lets you take advantage of parallelism when rendering a single QQL at high resolution.

How does it work? First, we compute layout information just once. Then, we divide the canvas into a grid: say, 2 cells wide by 4 cells high, to take advantage of 8 threads. Each thread performs a viewport-restricted render of just its grid. Once all the threads are done, we assemble their results onto the final canvas.

The main advantage is that each thread gets to skip over any circles that are entirely outside its grid cell. If there are many circles in the QQL that touch all the grid cells, then this will be less effective. However, even when few or no circles are able to be culled, we can still see significant speedups, which may be due to cache locality: each thread has a smaller amount of image data to work with. Even using a 2x1 chunk layout on QQL #26, where every circle needs to be painted in every chunk, I still see a 1.3× speedup! For QQLs that are better suited due to having circles well distributed, speedups on the order of 2–3× are more common.

Use the --chunks <WxH> option to specify a grid layout. For example, --chunks 2x4 uses a grid 2 chunks wide and 4 chunks high. Ideally, the product of W and H should be about as many threads as your CPU has. You can exceed that number (like, --chunks 16x16) but this will probably make things slower. The default value is --chunks 1x1, which corresponds to no parallelism.

In principle, you might see some seams or artifacts along the chunk boundaries. In practice, these tend to be entirely imperceptible, even with the aid of digital analysis. If you find otherwise, let me know! I'd love to take a look.

Incremental animations

TL;DR: Pass --animate points:100 to render a sequence of images as each 100 points are added, or --animate groups to render a sequence of images as each flow line group is added.

By default, rendering a QQL will create just one image: the final QQL. By passing the --animate <ANIMATE> flag, you can instead create a sequence of images that show the process of the QQL be drawn, circle by circle! The very first image will be blank, and each image after that will have more and more circles, until the last image, which is the complete QQL.

If you pass --animate points:N, where N is a positive integer (e.g., --animate points:100), each image will have up to N points more than the previous. If you pass --animate groups, each image will have one additional flow line group, which is a collection of points that were created together in the algorithm and have similar properties. Both modes, and especially --animate groups, can help you understand how the algorithm works.

Note that some points may be just barely off screen, or may have very small (or negative) radii. Therefore, some frames may be identical to their predecessors.

This example shows the flow line groups of QQL #53:

Incremental render of QQL #53, showing one group at a time.

When rendering for animation, the interpretation of the output path parameter (-o) changes a bit: frame numbers are added to each image before the file extension, zero-padded to 4 places. So, if you pass -o build/out.png, images will be created at build/out0000.png, build/out0001.png, and so on.

If you want to convert this image sequence to a video file for sharing, you can use FFmpeg. For example, if you rendered -o build/out.png and want to convert that to an MP4 video build/out.mp4, you can run:

ffmpeg -framerate 30 -i build/out%04d.png -vf format=yuv420p build/out.mp4

(The -vf format=yuv420p arguments instruct FFmpeg to use YUV 4:2:0 chroma subsampling; this slightly degrades color quality, but makes the video compatible with more players.)

Rendering with --animate is much faster than successively re-rendering QQLs. All the layout information only has to be computed once, and each frame only needs to incrementally render the new points. Most of the time goes to actually writing PNG files to disk. Exact speed of course depends on a lot of factors, but I typically see around 30 frames per second on my laptop.

You can also explicitly pass --animate none, which is the same as the default behavior.

Higher quality circles

TL;DR: If your --width is greater than 10000 or so, and you can see the straight lines in your circles, try adding --min-circle-steps 64.

In QQL, circles are approximated by polygons. The polygons for large circles have many segments, and the polygons for small circles have fewer segments (but never fewer than 8). At typical zoom levels, these just look like circles, but if you render at a really high resolution (say, 10k pixels wide at 100% zoom), you can start to see the individual line segments.

Use the --min-circle-steps <STEPS> option to increase the minimum number of segments. You should increase this approximately linearly with --width.

For example, here are two renderings of the "epicenter" of QQL #24 with a virtual canvas width of 40000px*. The first rendering leaves --min-circle-steps at its default value, while the second sets it to 64. Note that the circles are visibly smoother in the second:

default --min-circle-steps 64

If you can't spot the difference immediately, this visual diff may help you orient.

* Specifically, this shows a viewport of 0.030x0.030+0.650+0.651.

Fast collision checking

TL;DR: The --fast-collisions option may provide a moderate performance boost; it may also change the output.

Passing the --fast-collisions option changes the way that the algorithm checks whether two circles would collide. In fact, not only does it become faster, it also becomes more accurate! This sounds like a good thing, but it does mean that the answer to "do these circles collide?" can differ from the original algorithm, which can change the layout of the whole piece.

Radius inflation at paint time

TL;DR: The --inflate-draw-radius option dramatically changes the appearance of some QQLs that have lots of negative space.

When the QQL algorithm creates points along a flow field, each point gets a different radius. Sometimes a point gets a "negative" radius! These points still participate in layout, but don't get drawn at paint time, so QQLs with lots of negative-radius points tend to have… well, negative space.

The --inflate-draw-radius option keeps the layout step as is, but, at paint time, makes sure that each point has at least a small positive radius. This brings all the negative-radius points to life! Renderings affected by this can look "fuzzy" due to all the small points. Here's an example, using a seed from the marketplace:

Two versions of a QQL. At left: the QQL has a swirl in the top-right with a curious smattering of points around the rest of the canvas, but mostly empty. At right: all the negative space is filled in by colorful, silky tendrils moving along the flow field.

Important: Feel free to play around with this option, but if you share images with others, please make sure to communicate that these are not actually canonical QQLs!

Fidelity expectations

The output of this renderer is interpreted to match the canonical outputs "as well as the human eye can see at typical resolutions". This loosely defined goal is more lax than something like "bitwise identical to the pixel level", and there are some known deviations.

Canvas details

Both this implementation and the original use the stroke primitive on their respective canvas APIs to do all the actual painting. The implementations should emit "semantically the same" strokes, but the underlying canvas APIs may rasterize those strokes slightly differently. For example, you may see differences in antialiasing.

This purely affects painting, not layout, so should never change the overall appearance of a QQL. These kinds of differences tend to be visible only at very high resolutions (say, 10k pixels wide at 100% zoom) and when directly comparing the outputs of the two implementations.

Implementation-defined transcendentals

See docs on qql::rand::gauss for more details.

The original implementation uses Math.log as part of its Marsaglia algorithm for generating normal deviates. But this is defined in the ECMAScript spec as returning "an implementation-approximated Number value representing the result of the natural logarithm" (emphasis mine). The approximations may differ across implementations, so the original QQL implementation is not perfectly portable.

This implementation also uses the analogous f64::ln from the Rust standard library, which similarly has platform-dependent behavior. On my system, the results of f64::ln are consistent with the standard ln functions in C, Java, Python, Ruby, Haskell, …but not with Math.log in Firefox or Chrome's JavaScript implementations. The differences are very slight, on the order of 10^−16, and so this is unlikely to cause a visible change in the algorithm. However, if such a logarithm happened to cross a discontinuity—e.g., causing two points to just barely collide instead of just barely not collide—then this could in principle affect layout and cause visible large-scale differences.

(Similar for Math.sqrt, though implementations may be more consistent there.)

Since this behavior is implementation-defined to begin with, I don't consider such deviations to be bugs, but if you run into one I'd love to hear about it anyway, just for kicks.

Nondeterministic sorts

See docs on qql::rand::shuffle for more details.

The original implementation internally shuffles lists by pairing each element with a uniform deviate and sorting the pairs by the deviates. However, it doesn't account for the possibility that two deviates might collide, which happens with non-negigible probability since the uniform deviates are sampled from a space of size 2^32.

When two such deviates collide, the sort order under the original algorithm is implementation-defined, because the callback is not a "consistent comparator". In this case, the canonical QQL algorithm behaves differently in Chrome and Firefox, so there's really no "right answer". This algorithm will do "something reasonable" (i.e., produce a valid permutation of the input that was to be shuffled) but it may or may not correspond with either Chrome or Firefox's behavior.

As above, since this behavior is implementation-defined to begin with, I don't consider such deviations to be bugs, but if you run into one I'd love to hear about it anyway, just for kicks.

Intentional tweaks

Options like --fast-collisions and --inflate-draw-radius intentionally alter the behavior of the algorithm, in subtle or obvious ways. When these options are enabled, there may be significant deviations from the canonical algorithm; this is expected.

Bugs

If you're running into a layout or painting bug, and your QQL looks notably different when rendered under this implementation compared to the canonical one, let me know! Please send me the seed and the width that you're rendering it at, as well as a description of where in the image the difference is.

About

Rust-based QQL renderer with features and speed

Resources

License

Apache-2.0, MIT licenses found

Licenses found

Apache-2.0
LICENSE-APACHE
MIT
LICENSE-MIT

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages