COMPOSITION V3 / SAME SMALL NETWORK

Better fit. Composition still needs rules.

The retrained 40,994-parameter model predicts the held-out tile patterns much better than the previous weights. The procedural generator still preserves the focal view more reliably, produces fewer isolated wall fragments and runs faster. The model has not demonstrated better room design.

Same conditions. Different methods.

The primary comparison below uses the same held-out doors, purpose, landmark clearance and seeds. Both neural models use the same bounded connectivity sampler. Procedural geometry uses only the eight training grammars. The family-aware reference knows the held-out grammar and is shown separately as an authored target.

Tile NLL measures teacher-forced prediction, not generated-room quality. The composition proxy requires connected floors, protected clearance, a visible focal point, a clear discovery cell and wall density within a declared purpose band. It is not an enjoyment score.

The family breakdown matters.

The observation-room family regresses on focal visibility even though its tile prediction improves. Aggregate results should not hide that failure.

Inspect the compositions.

Two fixed examples per held-out family, chosen by dataset order. These were not selected for flattering results. L marks the entry view, F the focal landmark and D the optional discovery.