Blog / 5 min read · 2026-05-24

The trigger pattern — "When [action], [transformation]" for Gemini Omni VFX prompts

The syntax behind Gemini Omni's viral demos: mirror ripples, bubble sculptures, butterfly→bee morphs. How to structure trigger-based VFX prompts and lock the result.

The most viral Gemini Omni demos at I/O 2026 — the mirror-arm transformation, the bubble-to-glass sculpture, the butterfly morphing into a bee — share one prompt structure. DeepMind’s own guide names it the “trigger pattern” and it’s the cleanest way to get a controlled mid-clip transformation out of Omni.

The structure

[Base scene]. When [specific trigger action], [specific transformation]. Keep [list] identical.

Three parts:

1. Base scene — the starting state of the clip
2. Trigger — a specific action whose moment marks the transformation
3. Transformation + lock — what changes and what stays

The trigger is the syntactic anchor. Without it, Omni doesn’t know when in the 10-second clip the change should happen, and you get either a slow morph spanning the whole duration or a transformation jammed into the first second.

Worked examples

Example 1 — Mirror ripple (DeepMind’s official demo)

A woman reaches toward a mirror in a Victorian parlor.
When her fingertips touch the glass, make the mirror ripple like liquid
and her arm turn to reflective mirror material.
Keep the parlor, her face, and the warm lighting identical.

What works here:

• Base scene is concrete (Victorian parlor, woman, mirror)
• Trigger is a specific physical contact event, not a vague “after a moment”
• Transformation has two coordinated changes (ripple + arm material)
• Lock preserves three named elements

Example 2 — Bubble sculpture morph

A glass bubble floats slowly above a wooden table.
When the bubble reaches the top of the frame, transform it into a small
porcelain sculpture of a bird mid-flight.
Keep the wooden table, the soft window light, and the slow upward motion identical.

What works here:

• Trigger uses spatial position (top of frame) instead of physical contact
• Transformation is a specific named object (porcelain bird) not a vague morph
• Lock explicitly preserves the motion — the bird should still be drifting upward

Example 3 — Butterfly → Bee (Google launch demo variant)

A monarch butterfly hovers near a sunflower in a sunny meadow.
When the butterfly lands on the sunflower's center, transform it into a worker bee
collecting pollen.
Keep the sunflower, the meadow background, and the warm golden lighting identical.

What works here:

• Trigger is the landing moment — clear, single-frame
• Transformation keeps the same scale and intent (“collecting pollen” continues the action)
• Lock holds the environment so the focus stays on the morph

Trigger types Omni handles well

The pattern works because Omni has a clear before/after structure to plan around. Vague triggers force the model to guess timing, which produces either smeared morphs or abrupt jumps.

Trigger + lock + base scene — common mistakes

Mistake 1: missing the lock. Omni transforms but also re-styles the parlor / changes the lighting / swaps the subject. Fix: list 3-5 specific elements after “Keep [list] identical.”

Mistake 2: vague trigger. “When she touches the mirror” is weaker than “When her index finger touches the glass.” Be physical, be specific.

Mistake 3: trigger and transformation describe the same action. “When the leaves fall, the leaves turn red” — Omni doesn’t know which moment to lock to. Separate trigger from transformation: “When the wind picks up, the leaves turn red and start to fall.”

Mistake 4: trigger before generation. Triggers work during a clip, not as a setup. Don’t write “Trigger: butterfly lands. Show:…” — write a continuous prompt with all three sections inline.

Mistake 5: more than one transformation. Two coordinated changes (mirror ripple + arm) work. Three or more usually break — split into two turns using conversational editing instead.

Where to use trigger pattern

Common production use cases:

• Surface morphs — water → ice, fabric → stone, fire → ash
• Object swaps mid-frame — flower → fruit, glass → metal
• Character state changes — calm → startled, daylight age → night age
• Reveal animations — closed envelope → folded crane, closed flower → bloom
• Brand transformations — “When the logo lights up, the product behind it appears”

Avatar feature prompts can also use this pattern, but with @username as the subject:

@your_username sits at a desk reading.
When you look up at the camera, smile and wave.
Keep the desk, the warm lamp light, and your sweater identical.

This combines @username summon with a contained trigger — useful for talking-head intros.

When NOT to use the trigger pattern

• Static scenes — landscapes, ambient mood pieces. Add motion verbs (orbit, pan) instead.
• Continuous transformations — slow morphs spanning the whole clip work better as plain prompts (“watercolor painting slowly forming from blank canvas”).
• Multi-shot productions — for cuts between distinct shots, use Google Flow’s chaining, not triggers.

Quick checklist

Before sending a trigger-pattern prompt, verify:

• Is the base scene concrete (subject + action + location)?
• Is the trigger a specific, single-frame event?
• Is the transformation one or two coordinated changes (not three+)?
• Does the lock list 3-5 named elements to preserve?
• Is the whole prompt under 50 words?

Hit all five and your morph lands cleanly. Miss any and you usually get something weird in the rerolls.

Related

• Full field guide — covers the trigger pattern in context
• “Keep X identical” lock — the lock part of the trigger pattern, expanded
• VFX prompts in the library — production examples using this structure
• Glossary entry — formal definition