2bbf5bcf42
2 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Dejvino
|
2bbf5bcf42 |
Run both variety tests against the song bank, and fix the ceiling twice more
The tests now run on real bank entries. The seed gates used to run on a two-section synthetic whose only kinds were intro and outro, so half the scene library was unreachable and the number was measuring that rather than the seed. Probes are labelled by section kind and occurrence rather than by index, which is what makes two different songs comparable at all — they have different section counts, so probe 3 of one is not probe 3 of the other, and matching by position would compare a drop against an outro and score the mismatch as variety. For two seeds of one song the labels are identical and this changes nothing, which is the point. The song test measures one thing the seed test does not: coupling, the rank correlation between how different two tracks sound and how different their videos look. Separation alone can be had by a generator that ignores the audio and hashes the file, and that would be a perfect score for a completely wrong video. Separation without coupling is not variety, it is a different seed per file. The ceiling took two more attempts. Recasting every layer at random averages a dozen scenes together and a dozen random scenes converge on the same generic busy image, so two references came out closer to each other than two real videos and blocks scored over 100% of achievable. Forcing one scene per reference collapsed the other way: a video that never changes scene has almost no internal variation, so the ceiling landed BELOW the floor, which is a within-video quantity. A reference has to match the structure of what it bounds. They now keep the real pipeline — rosters, shots, per-section sampling, so a reference rotates between three or four scenes exactly as a real video does — while drawing from disjoint slices of the library. Same complexity, nothing in common. The fingerprint measurement had the same shape of error: pooling every probe mixed in how much each video varies over its own length, which is large for everything, and washed the answer to a flat 100% while the separation score said almost everything was collapsed. One vector per video now. Both tests fail as committed. Seed separation 0.04, song separation 0.03, coupling -0.03 — two different songs differ from each other by about as much as one video differs from itself, and that difference has no relationship to the music. Colour scores 112%. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Dejvino
|
656e062069 |
A seed variety test, and a metric that had to be gated before it was believed
Every other phase asks whether one video is correct. This asks whether two are different — the failure the suite could not see, since a generator that ignores its seed passes determinism, flash safety and liveness perfectly. Frames reduce to a structural descriptor built to be blind to the cheap axes and sensitive to the expensive ones: standardized luma kills exposure and palette, a Laplacian pyramid gives the radial spectrum, and the gradient-angle histogram is carried through a DFT magnitude so a rotation shifts it without moving it. Colour is measured and never counted; its only job is to expose the case where two seeds differ by a palette swap and nothing else. The score means nothing on its own, so it sits between two references the same instrument produced: a floor of how far one video travels from itself across its own sections, and a ceiling of the same pipeline with every layer recast at random. Two checks gate the instrument before any number from it is trusted — recolour must move structure ~0 while moving colour a lot, and a quarter turn must not move it at all. Three things this got wrong first and now does not. Averaging each video's probes into one descriptor washed out the structure being measured and put the floor above the ceiling; probes are matched instead, which is fair because the track is held fixed. Cosine distance on all-positive histograms scored unrelated scenes at 0.05, too compressed to be read; chi-square replaces it. Single-link clustering in the library sweep chained overlapping pairs into a fourteen-scene group that did not exist; complete-link means every pair inside a group is really a twin. The sweep runs against all 61 visualizations, and it finds what the per-scene distinct gate cannot, because that one compares raw pixels and structural twins are merely differently coloured. Both rendered gates fail as committed. That is the point of them: separation is 0.07 against a 0.35 target, two 4-cliques of scenes are one look each, and a third of the library is never cast. The instrument passes; the generator does not. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |