MODEL / IMAGE   —   LIVING LAB · V1

Watch the pieces
find their place.

Teach a small visual world. Break a new image into 36 fragments. Then watch learned spatial expectations and visual continuity pull the fragments into an arrangement. Nothing is moved using its original position.

V3 · Hierarchical sieve image model: learn, deform, remember and test real photographs →

NEW · Learn movable visual fragments and reconstruct what cannot be seen →

Experimental, deliberately readable.
Online memory + local relations + search.
No pretrained model. No network. No backprop required.
Source and tests ↗
Experiences taught0
Current test, placed correctly—
Pixel reconstruction error—
Search objective—

1. Give the system experience

Choose a repeating visual world. Each intact example updates 36 fragment prototypes and their neighboring feature differences. Tests use fresh examples that weren't taught.

Most recent experience · complete image
Learned slot memory · low-detail visual expectation

What changed inside?

These 36 locations store averaged *sieve features*: color core, a 3×3 luminance map, orientation and contrast. Neighboring locations also learn their expected feature differences. They are fixed spatial links, not automatically discovered object parts.

No experiences yet. Teach several examples and inspect the memory above.
Saved memory contains numerical prototypes, not photos. It stays in this browser.

2. Give it a puzzle it hasn't seen

All 36 image tiles remain available but arrive in random positions. The learner must find their arrangement. This tests assembly, not yet synthesis of unseen pixels.

Scrambled pieces · before inference
Held-out image (shown for your comparison, hidden from the solver)
The first test is fresh and scrambled. Give the system a few experiences, then press Assemble.

Inside the search

ReadEach 16×16 fragment → multiscale features.
RememberOnline location prototypes + adjacent differences.
SettleMatch tiles to slots, then minimize total learned and edge energy.
What this demonstration does — and what it doesn't

This is an interpretable assembly model. It learns from complete images without labels. A globally optimized one-to-one assignment and local swap search find plausible tile positions. The animated tile swaps show the resulting computed arrangement; they are not simulated neural firing. For evaluation only, the app retains original tile indices and measures exact placement and pixel error; the optimizer never reads them.

Its learned locations assume a stable visual layout. The built-in worlds are stylized and deliberately structured. For your own photos, teach multiple similar compositions, then test a different photograph. Diverse unrelated photographs, changing camera angles, repeated blank sky, and unfamiliar subjects can defeat this model. It does not yet invent details, discover movable object groupings, implement dendrites, or qualify as a competitive modern generative model.

Next research step: hierarchical coarse-to-fine groups with learned conditional predictions, explicit missing-patch synthesis, held-out real photographs, and controlled comparisons against seam-only jigsaws and ordinary image priors.