I generated 9,000 nonograms. Only 1 in 6 pictures survived 15 15.
I'm a junior solo developer from South Korea. I use a translator to help me write in English. My fourth app is Whale Squares: Nonogram.
In a nonogram, also called Picross, you use the numbers beside each row and column to fill in cells and make a picture. The first version had 90 puzzles. There were 30 each for 15x15, 20x20 and 25x25.
I'd heard that a nonogram app feels "a bit small" until it has around 10,000 puzzles. I thought my app needed something close to that to compete, so I aimed for 9,000.
Making three puzzles from one picture
I used image generation to make sheets of 16 pictures. Each picture was a black silhouette on a white background. I split the sheets into individual pictures and converted each one to all three grid sizes.
There were 9,680 candidate pictures in 21 batches.
I kept grids with 12% to 84% of the cells filled. A grid with too many empty rows or columns was out. Thin lines were a problem on a small grid, so at least half of the filled cells had to belong to a solid 2x2 block.
Then I ran a solver. It reads one row or column at a time. If the clues tell it that a cell has to be filled or has to be empty, it marks that cell. It goes through the rows and columns again until nothing changes.
If it can work out every cell this way, there is only one answer. Any cell it marks has to be the same in every solution. If there are still unknown cells when it stops, I leave that grid out.
In one batch, about 60% of the 15x15 pictures passed both the grid checks and the solver. At 20x20 and 25x25 it was about 98% for each size.
The message in a bottle looked like a fan
A puzzle can be solvable and have one answer, and still be a bad puzzle. If you finish it and can't tell what the picture is, there's no reward.
For the picture review, each finished grid had only a number beside it. The reviewer wrote down a first guess, then uncovered the name.
- A correct guess got 2 points.
- If the picture made sense after the name was shown, it got 1.
- If it still didn't make sense, it got 0 and that size was dropped.
The review was done on the same side that planned the sheets. I don't have guesses from independent players seeing the pictures for the first time.
At 15x15, a message in a bottle looked like a fan. A fish drying rack looked like two swings, and an oasis looked like one swing. In one batch of garden pictures, no flower passed at 15x15.
Take the bucket of water. On the small grid you still need to see the tilted bucket and the water falling out of it. Both parts tell you what the picture is. Some 15x15 grids were drawn cell by cell because the conversion lost a part like that.
Of the 9,680 pictures, 1,579 passed at 15x15 (about 16%). At 20x20, 4,612 passed (about 48%). At 25x25 it was 6,473 (about 67%). That's about 1 in 6 pictures on the smallest grid and 2 in 3 on the largest.
A picture could fail at 15x15 and still go into the game at the two bigger sizes.
I kept the 15x15 puzzles
After seeing the numbers, I thought maybe I shouldn't make 15x15 at all. In the end I kept that size and set the limit in the generator to 1,500.
The final pack has 1,500 puzzles at 15x15, 2,560 at 20x20 and 4,940 at 25x25. Of those 9,000 puzzles, 6,300 are pictures and 2,700 are patterns like clouds and waves. The patterns also have to be solvable by logic alone and have only one answer.
"One answer" didn't mean "good"
Later I got feedback that the puzzles were too easy. Of the first 30 puzzles at 15x15, 26 were left-right symmetric. Once you had the left half, you could copy it on the right. They were still solvable by logic, with one answer.
I added a rule for puzzles like these. I count a puzzle as nearly symmetric if 3% or fewer of the cells are different after a flip or a half turn. I also count puzzles where the solver can work out at least 90% of the cells on its first pass.
Together they can make up at most 10% of the puzzles in each size. I keep them out of the first 20% of each size's pack.
Solvable, one answer, and readable were not enough. The fourth check was simply: is this too easy?
Whale Squares: Nonogram on Google Play: https://play.google.com/store/apps/details?id=kr.binarytech.goraekan&utm_source=devto&utm_campaign=blog_202611
Whale Squares: Nonogram on the App Store: https://apps.apple.com/us/app/id6835263397
Comments
No comments yet. Start the discussion.