music-video-gen/flow-state/HOWTO-variety.md

How to build a visualizer that varies

HOWTO-visualizers.md covers making a scene that works. This one covers making a scene that looks different from one song to the next — a separate skill, and the one the library is currently worst at.

Everything here is a measurement rather than an opinion, and the numbers are quoted so a future change can contradict them. Several already contradict things believed earlier in the same week. Living document: add what you learn, and delete what stops being true.

---

The one-line version

A scene varies when the SONG can change what is on screen. It does not vary because its parameters moved.

Parameters mostly change the same picture. Content changes the picture.

---

How to know if you succeeded

gallery.html          your scene, six times, on six songs — with a score
checks.html?scene=X   the per-scene gate, including `consumes:` lines

The gallery score is the mean structural distance between a scene's own six frames, on the same descriptor the variety harness uses. Minimum bar: 0.04 (MIN_VARIETY in checks/gallery.js), drawn as a red line. Below it a scene is the same picture wherever it appears, and because it is cast into many songs that sameness leaks between them — the viewer recognises the shot rather than the song.

Some reference points, measured:

0.219  Droste Feedback      varies a lot
0.181  Spectrum Sculpture
0.049  Isometric Blocks     after work — see below
0.039  Apollonian Gasket    beautiful, and the same six times
0.031  Moiré Grid           the same picture in six palettes

---

What the descriptor can and cannot see

Design against this, because half of "why is my score low" is here.

It is blind to:

• Brightness and contrast. Frames are standardised to zero mean and unit variance before anything is measured. A scene that only gets brighter has not changed.
• Colour. Measured, reported, and excluded from the score. Six palettes cannot disguise one image, which was the entire point.
• Rotation, on the scale, orientation and texture blocks.
• Quality. It measures change between songs, nothing else. Apollonian Gasket scores 0.039 and looks great. Both facts are true and neither implies the other.

It sees:

block	what it measures	how to move it
scale	feature size — fine grain vs large forms	let the song set element size (stageScale())
orient	grid vs radial vs stripes, rotation-blind	break or vary a regular lattice
layout	where in the frame structure sits	move the subject, change what is empty
texture	element count, sparsity, mirror and radial symmetry	vary how many things there are
motion	what moves and where, not how much	vary the KIND of movement, not the rate

A known blind spot. layout is nearly useless for full-frame fields. Voronoi Shatter measures 0.004 there no matter what changes, because edge-to-edge cells occupy the same frame however they fall. That is honest — there genuinely is no arrangement — but it means a whole family is judged on four blocks instead of five. If your scene fills the frame corner to corner, expect to earn your score on orient and texture.

---

What has actually worked

Ordered by measured effect.

1. Draw the song's content instead of your own

The largest single lever. Take consumes: ['cast', 'ink', 'staging'] and draw castMain/castChorus where you would have drawn your own primitive. Recipe in MIGRATION.md.

Measured: identity is worth 117% of what the container is worth (checks.html?decompose=1) — swapping the song's cast under a fixed scene moves the picture more than swapping the scene under a fixed cast. A migrated library scene also carries the identity better than a stage written from scratch to carry it, which was a surprise and is why the whole library was migrated rather than replaced.

2. Let the song decide element SIZE

stageScale(). Not the size of your features relative to each other — the size of the whole vocabulary. A song of six huge forms and a song of four hundred tiny ones are different videos before anything else is decided.

Measured: adding elementScale to the identity moved stages from +0.0111 ±0.0043 to +0.0143 ±0.0008, and the scale block from 0.033 to 0.041.

3. Displace a regular lattice

If your scene is a grid, let the song push things off it. Bounded — past about half a cell the structure the grid was providing stops reading.

Measured: Isometric Blocks 0.034 → 0.049, with orient nearly doubling from 0.093 to 0.164 as the rigid lattice softened.

4. Give a full-frame field somewhere to be about

focusWarp(p, amount) and focusField(p). The song picks one to three focal points; tile in the warped coordinate so cells crowd toward them or pull away.

Measured: Voronoi Shatter 0.0463 → 0.0588. Note what moved: orient 0.140 → 0.194, while layout went 0.004 → 0.007, i.e. nowhere. Warping where cells sit changes what they look like without changing where the frame's energy is. The gain is real; the stated reason for it was wrong.

---

What has measurably NOT worked

Recorded because the failures were more informative than the wins, and because each of these looked obviously right beforehand.

Effects as parameters

Forty-nine scenes had their own glow; every one had its own grain. An effect knob sampled per song makes a scene look varied across parameter draws while its structure never moves — it inflates the score without changing the picture, and it fights the grade, which already does bloom, grain, chroma and vignette with an envelope the scene cannot see.

Do not add glow, bloom, grain, haze, chroma or trails to a scene. The post chain owns them. npm run lint:scenes does not catch this yet; reviewers should.

Concentrating disturbance instead of adding it

Isometric Blocks, twice. Spending the scatter budget near the focal points and calming the rest measured 0.049 → 0.0374 — most of the field went back onto the rigid lattice and took the orientation variety with it. Making the focus additive instead recovered nothing: 0.0365.

Two plausible diagnoses, both wrong. The lesson is narrow and worth keeping: a focal point has to be something the field gains, not something the rest of it pays for — and even that framing did not rescue it, so the real cause is still unknown.

Making the roster smaller

A whole afternoon went into the theory that a track drawing on fewer scenes would look more like itself. Swept directly across pool sizes 4, 8, 16 and 32 over twelve songs with three draws each: differences of 0.003 to 0.007 against a run-to-run noise of ±0.003 to ±0.005. Nothing. The theory came from an experiment whose arms differed in two ways at once.

Reading a form as a metric rather than drawing it

Spectrum Sculpture used sigShape as a radial distance rather than as a subject. Migrating it to castMain rendered pure black, because the cast carries notches and a hollow and an annulus used as a radius turns a sculpture inside out. If your scene consults a form's geometry rather than drawing it, it wants sigShape, and the cast is not for you.

---

Scene shapes, and how hard each is to vary

• Element-placing (metaballs, floating-geometry, firefly-drift) — easiest. Take the cast, place on stageNode, keep your motion. Nearly all the wins above are this shape.
• Single figure (slow-orb, eclipse-field) — take the cast and stageScale(). Composition is yours to vary; most such scenes centre their subject and never move it, which is free layout left on the table.
• Full-frame field (voronoi-shatter, plasma-bloom, curl-flow) — hard. layout is unavailable to you. Earn it on orient and texture, and consider a focus.
• Fixed geometry (truchet-fold, quasicrystal, apollonian-gasket) — hardest, and currently unsolved. The image is the maths, so the cast cannot be pasted on top of it. The open idea is to make the cast the repeating UNIT the tiling is built from, which is a real rewrite per scene rather than a recipe step. Nobody has tried it yet.
• Treatment (analog-wow, scan-tear, halftone-misprint) — these are effects wearing a scene's clothes. They score 0.031–0.040 and their whole migration was one posterisation. They probably belong in the identity's effects register rather than competing for screen time as subjects.

---

Trusting the numbers

This harness has two classes of measurement and only one is safe to steer by.

Direct render comparisons — the gallery score, checks.html?decompose=1, the per-scene gate — compare rendered frames to each other. Noise floor of exactly zero. Trust these.

Aggregate ratios — chiefly separation, which divides one small difference by another — swing wildly. Three runs of the same post-migration measurement gave 0.31, 0.56 and 0.07. Four conclusions were drawn and withdrawn during this epic for exactly that reason. Anything under about 0.01 of spread needs &repeats=3 before it is believed, and a difference that changes sign with the sample size is not a difference.

Two more traps, both of which have bitten:

• A permanently-zero block looks like a property of your scene. The gallery built four of the descriptor's five blocks for weeks; motion came back 0.000 for every scene and depressed every score by a fifth. If a block is identical across many unrelated scenes, suspect the instrument.
• Arms that differ in more than one way cannot be read as if they differed in one. This produced the roster-size result above, and it was believed for a day.

---

Checklist for a new or reworked scene

1. consumes declared, and the consumes: lines pass in checks.html?scene=<name>. A declared artifact that is ignored is worse than one not declared, because the casting code believes it.
2. No glow, bloom, grain, haze, chroma or trail parameters.
3. Element size goes through stageScale().
4. Placement goes through stageNode() unless the composition IS the scene.
5. Gallery score above 0.04, checked before and after.
6. The rest of the per-scene battery still passes — especially distinct. The more scenes share a cast, the easier it is to become one of your neighbours.

---

Open questions

• Coupling is zero and has never moved. Whether a song looks different in proportion to how it sounds different measures -0.02 ±0.03. Every change in this epic left it there. Nobody has attacked it directly.
• Fixed-geometry scenes have no known route to variety.
• The layout blindness for full-frame fields would need a new descriptor block measuring cell statistics rather than where structure sits.
• The 0.04 bar is a starting point, set where the flat cluster sat — and set against scores that were depressed by the motion bug. It deserves re-deriving.