Synthesizes a turtle program that redraws a black-and-white image. MCMC search over a tiny DSL — no neural net. The render improves the longer it runs, converging toward the target.
- Load the target PNG, resize to 256×256, binarize (on if luma < 128).
- Search for the turtle program whose rendering best overlaps the target.
- Save the binarized target and the best result.
A program is a fixed vector of 512 commands, each carrying two u8 params:
Forward (stroke), Turn/SetAngle (heading), MoveTo (jump), SetWidth, SetColor,
NoOp. Running the program draws strokes onto a 256×256 1-bit canvas; fitness is the Jaccard
index (overlap / union) between that render and the target.
Search is MCMC with an exponentially cooled temperature: each step proposes a mutation of one command (70% small parameter tweak, 30% full random replacement) and accepts it via the Metropolis rule. Several independent chains run in parallel — best chain wins.
cargo build --release
# morph <image> [time_secs=60] [chains=1]
target/release/morph kiwi.png 60 4Writes kiwi_target.png (binarized target) and kiwi_result.png (best rendering).
Set MORPH_CHECKPOINT_SECS="60,600,1200" to dump best-so-far snapshots at those times.
See scripts/ for batch and checkpoint runners.
4 chains, 3h (10800s) budget
| Target | Result | Jaccard |
|---|---|---|
 |
 |
0.976 |
 |
 |
0.956 |
 |
 |
0.974 |
 |
 |
0.983 |