Don't regenerate the clip — just say what to change in it

Here's the one-line idea: you've got a short clip — an animated photo of your street, a five-second demo of your project. You want it to be night, not day. You type one phrase: "make it night." And only the sky changes. The cat, the angle, the camera move — all as they were.

This isn't "build a clip from one phrase" — that one already works. This is about editing something you already made: change one detail without losing the whole frame.

Why this just became possible

Editing AI video used to be blunt: don't like a detail — generate the whole thing again. And you get a different clip. Different cat, different light, different motion. Fix one thing, break ten.

On June 30, 2026 Google opened Gemini Omni Flash in the API — a model that edits video by conversation. The key word is "remembers." Straight from Google's docs: "the model remembers the video context, applying your changes while preserving elements you did not mention." Each edit references the prior result. Say "add a slow zoom" and it adds a zoom, it doesn't reshoot the scene.

What you'll learn

Small project, but you touch the useful stuff right away:

• Call a model over an API — the same Gemini key. Send an idea → get a clip.
• Keep state across turns — each edit references the prior response (previous_interaction_id). That's how the model "remembers" the clip.
• Animate a photo — the input can be a still, not text, and it starts moving.
• Count by the second — video runs about $0.10 per second. Good to feel the cost of a request.

A ready starter prompt

The trick isn't the first prompt — it's the edit. A bad edit asks to "redo it" and you lose the frame. A good one names one thing and forbids touching the rest.

The weak version is a gamble: the model rebuilds the scene its own way. The strong one sets boundaries, and only what you named changes.

What you'll end up with

You had a daytime photo of your street. Now it's a five-second clip with a slow zoom across it — and on the second edit, sunset. Same house, same angle, but cinematic. And you built it with phrases, not in an editor.

Honest about the limits: clips are short (3–10 seconds), quality is preview-grade, and each one carries an invisible SynthID mark that it's AI. For a weekend project, exactly right.

Weekend plan

1. Friday night. Grab a Gemini key, open AI Studio, and make one clip from text. Then change a detail in it with one phrase. That's the "whoa" moment.
2. Saturday. Wrap it in a tiny app: a "make a clip" field + a "what to change" field. The button sends an edit that references the prior response.
3. Sunday. Let it animate your photo, and save the edit history so you can roll back.

Start with one edit on one clip. A chain of five changes comes later — first, get "change one thing, everything else stays" working.

Source: Google AI for Developers — Gemini Omni Flash