Gemini Omni Flash is two tools in one prompt box: it generates video from text or an image, and — uniquely in Media Forge — it edits existing footage from a plain instruction. Generation prompts work best as short narrative scene descriptions, not keyword lists. Edit prompts follow a different grammar: name one change, then fence everything that must stay untouched ("make it night; keep the camera move, framing and subjects identical"). There are no extra dials — no resolution picker, no audio toggle, no seed — so the prompt does all the steering.
One flowing paragraph describing the scene the way you would to a cinematographer — subject, place, moment. Gemini models respond to narrative description far better than to comma-separated keywords.
How the shot moves and how the world moves: "slow drone glide", "static frame while the crowd flows past". One motion idea per clip — these are short generations (3-10s), not montages.
Photographic light language: "golden hour side light", "overcast soft light", "neon-lit night". Time of day is the single strongest lever on the look.
One transformation, named plainly: "make it night", "turn the car red", "make it snow". The best edit instructions read like a note to a colorist, not a prompt.
One sentence fencing what must survive: "Keep the camera movement, framing, subjects and timing exactly the same." Without it the model may reinterpret more than you asked.
Template
Generation: [subject and action] in [specific setting], [camera or world motion], [lighting and time of day], [mood or style].
Edit: [one plain-language change]. Keep [camera movement / framing / subjects / timing] exactly the same.A drone glides slowly over a coastline village at golden hour, whitewashed houses stacked on the hillside, fishing boats leaving thin wakes in the harbor below, warm low sun flaring gently across the lens, calm and cinematic.
A single flowing sentence with one camera motion, a concrete place and a time of day — the shape Omni Flash likes. "Golden hour" does more grading work than any list of quality adjectives would.
Close-up of matcha powder sifting down into a white ceramic bowl in soft slow motion, a bamboo whisk resting beside it, bright natural daylight from a window, minimalist cafe aesthetic, gentle and satisfying.
Short durations are this model's comfort zone: a 3-4 second single-action clip loops cleanly on social. One action, one surface, one light source — resist adding a second beat.
A static camera holds on a rainy Tokyo crossing at night as umbrellas stream in every direction, neon signs reflecting in the wet asphalt, one person in a yellow raincoat standing still in the middle of the flow, moody and cinematic.
Flipping the motion — static frame, moving world — is a reliable recipe here. The single yellow raincoat gives the eye an anchor; name one focal detail in any crowd scene.
Bring this product photo to life: the camera pushes in very slowly toward the bottle while soft steam drifts through the background light, the label stays perfectly sharp and unchanged, everything else in the frame holds still.
In image-to-video the still supplies composition and branding; the prompt adds only motion. Explicitly protecting the label ("stays perfectly sharp and unchanged") is the same scope-guard habit that pays off in edit mode.
Animate this portrait subtly: hair moves in a light breeze, eyes blink naturally, the fabric of the collar shifts slightly, background bokeh shimmers. The pose, framing and expression stay as in the photo.
The cinemagraph recipe — secondary motion only, pose locked. Small-motion prompts preserve the source image's fidelity; asking a still portrait to turn or walk is where i2v breaks.
Make it night. Streetlights and window lights come on, the sky turns deep blue-black, reflections appear on the road. Keep the camera movement, framing, subjects and timing exactly the same.
The signature Omni Flash edit: a global relight of real footage from one sentence. Describing the consequences of the change (lights on, reflections) guides the re-render; the scope guard keeps the shot itself intact.
Turn the car red — a deep glossy red, with correct reflections on the paint. Change nothing else: the driver, the street, the lighting, the camera motion and all other vehicles stay exactly as they are.
Object-level edits work best on one clearly identifiable subject. Naming the finish ("deep glossy red, correct reflections") steers the material, and the explicit "all other vehicles" line prevents collateral recolors.
Make it rain: steady rainfall, wet reflective pavement, drops streaking through the light, subjects unchanged and dry under the awning. Keep the framing, camera movement and timing identical to the original.
Weather is a whole-frame transformation, which is exactly what this editor is strongest at. Small physical consequences in the instruction ("wet reflective pavement") make the rain sit in the scene instead of on top of it.
Regrade this clip with a cinematic teal-and-orange look: warm skin tones, teal shadows, gentle contrast lift, slight film softness. Do not change any content — same subjects, same motion, same framing, same timing.
Used this way, Omni Flash is a prompt-driven color suite for footage you shot or generated elsewhere. Keep grade language photographic (tones, shadows, contrast) rather than emotional ("moodier") for repeatable results.
Turn this summer scene into deep winter: snow on the ground and rooftops, bare trees, visible breath from the people, cold blue-grey daylight. The people, their clothing colors, the camera path and the timing stay the same.
Season swaps combine relight, set-dressing and atmosphere in one instruction — a showcase edit for before/after content. Decide what the guard protects (here: people and camera) and say it, because "everything" is not specific enough for a change this large.
Re-render this clip as a hand-painted watercolor animation: soft brushstroke edges, paper texture, muted pastel palette. Preserve the exact motion, composition and timing of the original footage.
Full style transfer on real footage — the motion survives, the medium changes. Committed, physical style descriptions ("brushstroke edges, paper texture") transfer better than art-movement name-drops.
Pass 1: Make it night, streetlights on, keep everything else identical. — Pass 2 (run on the result of pass 1): Now regrade it with a cool cinematic blue tone and light film grain, changing nothing else.
The conversational workflow: one change per pass, each edit run on the previous output. Two clean passes beat one overloaded instruction — stacked changes in a single prompt are where edits start reinterpreting the shot.
Whole seconds only, 3-10 (the picker offers 3, 4, 5, 6, 8, 10). Use 3-5s for loops and single actions, 8-10s for scenes with a camera move. Edit mode takes no duration — you send only the clip and the instruction.
16:9 for landscape, 9:16 for vertical social — the only two options, so plan the framing for one of them. Edit mode works on the clip you upload; ratio is not part of the instruction.
Text-to-video for new shots, image-to-video to animate a still, edit mode to transform existing footage. Edit is the mode no other Media Forge model has — reach for it before regenerating a shot you already like.
There is nothing else — no resolution picker, no audio toggle, no seed. The prompt is the entire control surface, which is exactly why the scope-guard sentence matters so much in edit mode.
Omni Flash inherits the Gemini family trait: a narrative paragraph beats a tag list every time. "Cinematic, 4k, golden hour, drone, coastline, beautiful" produces generic footage; "a drone glides slowly over a coastline village at golden hour, fishing boats leaving thin wakes below" produces the shot. Put the strongest levers in every prompt — a concrete place, a single motion idea, and a time of day — and treat adjectives as seasoning, not structure.
Every reliable edit instruction has two halves. The change: one transformation stated plainly, with its physical consequences ("make it night — streetlights on, reflections on the wet road"). The guard: an explicit list of what survives ("keep the camera movement, framing, subjects and timing exactly the same"). The editor re-renders the whole clip to apply your change, so anything you do not fence is open to reinterpretation. The guard sentence is not boilerplate — tailor it to the shot: protect the label on a product clip, the faces in a people clip, the camera path in a move.
Treat edit mode as a conversation with your footage: run one instruction, review the result, then run the next instruction on that output. Night pass, then grade pass, then weather pass — each edit compounds on the last while staying individually reversible, since every intermediate clip stays in your library. This also isolates failures: when a three-change instruction goes wrong you learn nothing, but when pass two of three goes wrong you know exactly which instruction to reword. Finished chains drop straight onto a Media Studio timeline for sound and titles.
Edit mode. Every video model in Media Forge generates; Omni Flash is the only one that takes footage you already have and re-renders it changed — "make it night", "turn the car red", a full regrade or restyle — from a plain-language instruction, keeping the original motion and framing.
Any clip you upload in Media Forge — footage from another model, your own camera footage, or a previous Omni Flash output. Edit mode takes just the clip and the instruction; there is no duration or ratio to set.
There is no audio option on this model, so don't spend prompt space directing sound. Treat the output as picture-first and add music, voiceover and sound effects in Media Studio, where you control the mix anyway.
Generation runs 3 to 10 seconds in whole seconds (3, 4, 5, 6, 8 or 10 in the picker), in 16:9 or 9:16. That is the full option surface — there is no resolution or seed setting, and edit mode needs only the source clip and the instruction.
You can ask, but you shouldn't — stacked changes in one instruction are the main cause of the editor reinterpreting the shot. Run one change per pass and feed each result into the next edit. Two clean passes are faster than re-rolling one overloaded instruction.
The instruction had no scope guard. The editor re-renders the entire clip to apply a change, so end every edit prompt by fencing what must stay identical — camera movement, framing, subjects, timing, and any detail you care about, like a product label or a face.