Word Dials
A dial from four words, with no prompt pairs.
Every other dial on this site was made from about ten pairs of prompts. These were made from a list of four words, like underwater, submerged, bubbles, aquatic, by asking the text encoder which direction makes it more disposed to say those words. This is part of ongoing work with fredbliss, who brought the Jacobian lens to H3 and is trying keyword dials with it; what follows is our own reimplementation, and we compare notes as we go. Same scene, seed and strength for every clip; Pika H3 Flex, 6 steps, text-to-video. Turn the sound on: half of the result is in the soundtrack.
read this firstThe control and a placebo
A dial this strong reshuffles any scene a little, whatever it means. So every clip below is judged against a dial of the same size made from four bland words (table, number, between, also). The placebo gives a different train and a quieter soundtrack (−9 dB), and nothing else.
Train platform at dusk, seed 42
worked on the first tryUnderwater and thunder
Underwater turns the platform into a harbourside: open water, moored boats, and the train ploughing through water with a bow splash. Thunder is mostly heard: the soundtrack is ~12 dB louder than the control (~21 dB louder than the placebo, which quiets things) and its energy below 150 Hz doubles, with windswept hair and a low, dramatic sun. Its nearest prompt-pair dials in our library are the storm and voice-energy dials; underwater's nearest include our underwater dial.
Four words each, strength +2
a surpriseWords become places
Words with a strong sense of place change the setting rather than the surface. Stained glass (stained, glass, cathedral, mosaic) builds a grand station hall under a vaulted glass roof, not coloured glass. Origami moves the whole scene to a Japanese station: the dial carries what the word evokes, not folded paper.
Place words, +2
two rowsCombinations
Two word dials at 1.5 each. Thunder survives being mixed with underwater (stormy, steam blowing off the train, +6 dB), but the water does not appear at this strength. Stained glass + hush gives the glass hall again.
Two rows at 1.5 each
honest resultsWhat did not work
Words for a surface look mostly did not come through: bioluminescent (no glow), static (a colder blue hour, no TV static), vertigo (a different station, little motion). And the shortcut failed completely: reading words straight off the model's output layer (a "logit lens") gave dials that behave exactly like the placebo, underwater included.
Look words, +2
The shortcut: logit-lens word dials, +2
methodHow it works
- H3's text encoder is the first 50 layers of Qwen3-VL-32B-Instruct. We checked that its weights match the upstream model bit for bit, then borrowed upstream's last 14 layers and output layer, which H3 does not ship.
- For each word, the Jacobian lens gives the direction at H3's encoder output that makes the model more disposed to say that word, averaged over 56 prompts (paper). A list's direction is the mean over its words.
- That direction is pushed through H3's own projection and token refiner, exactly where Make Dial measures a prompt pair, and the result is debiased and scaled to the size of a typical dial.
- The key step: every word list shares one big common direction, the bland list included. Only after removing it do the dials point at their meaning.
One scene and one seed per clip, so these are leads, not results. The word dials are not released yet. Ideas, plan and numbers live in the Dials research notes.