Learned gallery
Unfreeze a compiled program’s weights & train them in real time.
- Drag a rectangle
- Drag its corner to resize it
- ClickTap empty space to add one, up to 12
- Remove the selected one with Delete or the Delete button
More detail
What’s happening
A tiny neural net places the photos. It has adjustable numbers, called weights, and every photo’s position and size is computed from them. When you edit the sketch, the page nudges the weights about times a second until the photos land where you drew them. That’s training, and it runs live in this tab.
The loss
Training needs a score to shrink, called the loss. The chart draws its two parts: how far the photos are from the sketch, in copper, and how much edges that nearly line up miss each other, in indigo. Every step changes the weights a little so that the score drops. Pairs of edges that stay apart get pulled harder over time, so they end exactly on one line.
The slider
The misalignment penalty sets how tidy the result is. At 4%, two edges of one kind, e.g. two tops, that are closer than 2% of the stage’s width end on one line, even when the sketch is a bit off. Slide it to off, and the photos simply copy the sketch.