Board Size Chart
| Board Size | Levels | Images | Order |
| 70×70 | 18 | 32 | 4 |
| 71×71 | 13 | 20 | 2 |
| 72×72 | 16 | 22 | 3 |
| 73×73 | 14 | 16 | 1 |
A total of 48,000 levels were generated with 12,000 of each size.
I included the order they were generated in because this influenced the distribution of levels/images, with the more recent batches having an advantage.
It’s About The Colors
I made an effort to select images that have relatively unusual colors compared to my typical choices. This means a lot fewer images with red/yellow highlight colors.
I believe that my recent natural line update ,particularly the increased line glow, makes certain color combinations work that didn’t work before.
The lines, on the other hand, were relatively less important in terms of my selection process. I was OK with including patterns that aren’t particularly special if the colors justified it.
Yellow Similarity Shading Is Back
I stopped using yellow (green/red) similarity shading after the 50×50 set (a long time ago) because I generally preferred images with magenta(red/blue) and cyan (green/blue) shading.
Magenta shading dominated most of the images since, and The 80s Unedited was no exception – 40 out of the 45 images had magenta shading.
There are 46 magenta shaded, 32 cyan shaded and 12 yellow shaded images in this set.
The increased line glow as a result of the natural line update makes it harder to tell which is which – it is most noticeable around the edges of the board. Here is one of the more easily recognizable yellow shaded images:
Almost Unedited
While I can’t claim the set is 100% unedited like the The 80s Unedited set, the handful of line edits I made were in the first two batches (73×73 and 71×71). In total I edited the lines of about 10 of the images, changing the color of no more than 4 lines. Almost all of these were bright short lines that I didn’t feel fit with the rest of the image.
The Program Determined the Image Order
The order of the images is based on an algorithm I wrote designed to maximize the difference between adjacent images in the set primarily based on the colors.
I spent far too long on the implementation of this process:
besides collecting data based on the pixel values of each image, I wrote a separate app that randomly showed me two images and I rated them based on similarity. After doing this for about 500 pairs of images, I used linear regression to generate a formula that predicted how similar I would rate them myself.
This was then used to feed yet another app I had to write that maximizes the score of images near each other in the set while keeping each levels images near each other.
The work might pay off if it can be used to programmatically filter out images without my having to look at them, but I think the process needs to be improved significantly before I consider this.
Previous gallery: The 80s Unedited