MODELIMAGE · EXPERIMENT THREE · READABLE GENERATION

Let a world teach
its own visual structure.

A multi-frequency sieve, small competitive populations and movable relationship hypotheses. Train it on whole scenes, conceal part of a new one and watch its remembered structure propose the missing pixels.

1. Give it experiences

Every intact image updates three competitively learned visual dictionaries and the relations between them. Group candidates compete on how well they predict other fragments, rather than using semantic labels.

0Complete experiences
0Spatial one-part rules
0Candidate group relations
No experiences. All three dictionaries start empty. Use a synthetic world first, or select several of your own photos.

The three learned frequency populations

Coarse (2×2 per patch), middle residue (4×4), and fine residue (8×8). The synthesis reconstructs their sum; each prototype is inspectable and changes under competitive Hebbian adaptation.

0Coarse atoms
0Middle atoms
0Fine atoms

Coarse prototypes

Middle residues

Fine residues

3. Measure, don't guess

When testing photographs, use images the learner has never been taught. Here you can choose separate test files. The benchmark compares pixel reconstruction error on hidden regions only.

Select unseen test photographs. The repo also includes a reproducible 12-image-crop benchmark from four completely held-out scikit-image photos; its CI measurements are reported separately from this browser session.

2. Ask it to fill a missing piece

Actual image · evaluation only
Visible sensor pixels · missing region blocked
Nonlearned spatial-color interpolation
Learned coarse reconstruction
One-part context + three-level synthesis
Learned groups + coarse→middle→fine synthesis
Trust-checked reconstruction · memory accepted or rejected

Which learned relationship did it use?

Click a hidden position on the sensor image to inspect its chosen explanation. Solid turquoise lines show fragments the predictor used. If nothing meaningful has been learned, it will fall back to a generic atom. The separate trust-checked image tests whether the learned hypothesis is reliable on still-visible pixels, and can suppress it.

No relation selected yet.
Mean squared pixel error is reported only inside the hidden mask. Lower is better. Visually attractive hallucinations can still score worse than simple interpolation.
Architecture and honest scope

This is a small interpretable generative completion model, not a diffusion model or proof of conscious experience. Three scale-specific dictionaries learn through winner-take-all normalized Hebbian-style updates. The full 8×8 patch decomposes exactly into a 2×2 base, 4×4 intermediate residue and 8×8 fine residue. Generating a hidden patch uses a learned coarse code, the most frequent matching conditional middle code, and a conditional fine residue.

Spatial templates are shared across all 12×12 patch positions and test candidate one- and two-part relations. One-patch geometric slack allows a limited form of deformation; this is not a learned continuous warp or a semantic object detector. Candidates win on predictive lift and support, not labels. Uploaded images remain in your browser and are not stored remotely.

The hard problems remain: real-world scale/viewpoint invariance, discovery of higher-order objects without combinatorial explosion, meaningful conditional high-frequency generation, and consistent results against interpolation and convolutional baselines. The included held-out real-photo benchmark may reveal negative transfer. Failed gates are still results.