Procgen Tarot

Been poking around all day, and now I have some coherent thoughts on improving the image generation.

The pips have some good variety, but at the moment aren’t different enough that, e.g., diamonds and chimes can be told apart by shape in a given deck. 

An easy way to vary them more would be to alter their “density” or starting number of seeds. I’ve cranked up the seeds and gotten an interesting look.

Very occasionally, I get a suit whose pips favor some zigzag lines. I got a “rocks” suit on one deck which was a grey color and did this! It looked like jagged boulders. Possibly this outcome should be slightly more common to increase variety.

I suppose mask shapes could also be varied more. Thin, randomly tilted rectangles, crescent shapes, or annuli (circles with a hole) might look noticeably different, but I haven’t tried any of that.

A more out-there idea would be to add a “foreshortening” effect to either the line quality or the flow fields. Right now, flow fields have “right” and “left” which are at right angles from their forward direction. But these “right” and “left” vectors could be placed at shallower angles to the forward vector, with “right” still being the negative of “left”. Applying foreshortening could potentially make some objects seem more 3D. And, flow fields could potentially have different amounts of foreshortening in different regions. (At this point it’s almost like having two different flow fields, one conceptualized as at a right angle to the other.) This means they could represent the geometry of foreshortened curved surfaces.

Besides the pips, the other major thing to my mind is the figures. Sometimes they turn out very interesting, but they have some quirks which can feel repetitive; such as not having a clear head or being roughly diamond-shaped. What I’ve done for myself is add some extra mask shapes, such as just the top half of an ellipse, and a plain rectangle. The plain rectangle doesn’t look too great, but maybe more random quadrilaterals would have some character. I was also thinking maybe a more hand-shaped human silhouette could be good, or a selection of appropriate Julia sets.

I’d also like to try and see what happens when all the density is placed near the edges of the mask shape, leaving the middle basically empty. But another option would be to try distributing the density by Perlin noise.

When it comes to having a head, probably the only solution is to put some sort of deliberately generating head near the top of figures some of the time. Wouldn’t want to do that too much. The strength of this comes when it enables seeing things in the scribbles, not when it forces one interpretation.

If you don't mind, here's my current understanding of how your image generation works - how close am I to correct?

Everything sounds about right. I'm not sure about fine details as this generator was made more than two years ago.

One thing I'm not clear on, is how many growth points initially are placed onto the card. 

As far as I remember, a single tracker is placed per time (or something like this). I tried to spawn all the trackers according to the density, but it looked bad: this way they didn't have chances to draw long curves because they bumped into each other too soon.

As I understand it, growth points actually spread outward in two directions (unlike Substrate).

To be honest I have no idea what Substrate is. Yes, curves grow in two directions. Otherwise there would be loose ends which could cause problems with building of closed contours.

I will add svg export: these are vector images, so it shouldn't be too hard.