Gemini Omni Flash

Introduction

Text-to-video gives you the model's interpretation. The "young woman in a cream linen shirt" is whichever face the weights settle on. The "Parisian cafe terrace" is whichever cafe the model imagines from millions of training images. Fine for exploration, wrong for brand work, recurring talent, specific products, or a hero shot the marketing team already signed off on.

Gemini Omni Flash takes the opposite approach when you pass inputs.referenceImages. You hand it up to seven images describing exactly the style, characters, or locations the output should honor, and the model uses them as visual ground truth rather than as prompt suggestions. Google's prompt guide names three distinct uses for this workflow: style transfer, character consistency, and storyboard guidance. This guide covers all three plus the combined first-frame anchor mode.

The reel above is one API call with four reference images: the traveler shown below, plus three location stills the model treats as the truth for each city. The traveler reads as the same person in all three cities, and each city matches its reference exactly: the same Parisian cafe terrace, the same Tokyo backstreet at dusk, the same Marrakech rooftop at golden hour.

A modern female traveler in her late twenties, light brown wavy shoulder-length hair pulled back loosely, fair skin with a light tan, warm friendly expression, cream linen button-down shirt with sleeves rolled to the elbows, a small black analogue camera on a leather strap around her neck, a tan leather messenger bag at her shoulder

A sunlit Parisian sidewalk cafe terrace in the morning, framed as a clean square composition. Small round white marble tables with cane wicker chairs, espresso cups and a small white pitcher, a folded newspaper, a cream cafe awning above, a Haussmann building facade with wrought-iron balconies behind, a vintage bicycle leaning against a lamppost on one side

A vibrant Tokyo backstreet at dusk, framed as a clean square composition. Tall narrow buildings on both sides covered in colorful Japanese neon signs in pink, cyan, and yellow. Paper lanterns hanging across the alley above, wet pavement softly reflecting the neon

A Marrakech rooftop terrace at golden hour, framed as a clean square composition. A low handwoven Berber rug spread across terracotta tile, two large embroidered cushions in deep reds and oranges, a silver Moroccan tea set on a low brass tray. The Koutoubia minaret rises behind, silhouetted against the warm pink-orange sunset sky

The rest of this guide walks through each of the reference workflows in isolation, then closes on the combined frame-anchor mode for when the opening frame also has to be locked.

Request shape

A reference-driven Gemini Omni Flash request takes a prompt and inputs.referenceImages. Dimensions are optional and resolve to 720p in either 16:9 or 9:16:

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'google:gemini@omni-flash',
  positivePrompt: '...the traveler from the first reference visits the three cities in the order of the next references...',
  inputs: {
    referenceImages: [
      'https://example.com/traveler.jpg',
      'https://example.com/paris.jpg',
      'https://example.com/tokyo.jpg',
      'https://example.com/marrakech.jpg'
    ]
  },
  width: 1280,
  height: 720,
  duration: 8
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "google:gemini@omni-flash",
            "positivePrompt": "...the traveler from the first reference visits the three cities in the order of the next references...",
            "inputs": {
                "referenceImages": [
                    "https://example.com/traveler.jpg",
                    "https://example.com/paris.jpg",
                    "https://example.com/tokyo.jpg",
                    "https://example.com/marrakech.jpg"
                ]
            },
            "width": 1280,
            "height": 720,
            "duration": 8
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "model": "google:gemini@omni-flash",
      "positivePrompt": "...the traveler from the first reference visits the three cities in the order of the next references...",
      "inputs": {
        "referenceImages": [
          "https://example.com/traveler.jpg",
          "https://example.com/paris.jpg",
          "https://example.com/tokyo.jpg",
          "https://example.com/marrakech.jpg"
        ]
      },
      "width": 1280,
      "height": 720,
      "duration": 8
    }
  ]'

runware run google:gemini@omni-flash \
  positivePrompt="...the traveler from the first reference visits the three cities in the order of the next references..." \
  inputs.referenceImages.0=https://example.com/traveler.jpg \
  inputs.referenceImages.1=https://example.com/paris.jpg \
  inputs.referenceImages.2=https://example.com/tokyo.jpg \
  inputs.referenceImages.3=https://example.com/marrakech.jpg \
  width=1280 \
  height=720 \
  duration=8

{
  "taskType": "videoInference",
  "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "model": "google:gemini@omni-flash",
  "positivePrompt": "...the traveler from the first reference visits the three cities in the order of the next references...",
  "inputs": {
    "referenceImages": [
      "https://example.com/traveler.jpg",
      "https://example.com/paris.jpg",
      "https://example.com/tokyo.jpg",
      "https://example.com/marrakech.jpg"
    ]
  },
  "width": 1280,
  "height": 720,
  "duration": 8
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "videoUUID": "9c1b2d3a-4e5f-6789-abcd-ef0123456789",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/9c1b2d3a-4e5f-6789-abcd-ef0123456789.mp4"
  }
]

One required field, the rest depend on the mode:

• positivePrompt is required, minimum two characters. Google's prompt guide emphasizes that Omni Flash is less prescriptive than Veo, so short focused prompts usually outperform long over-instructed ones.
• inputs.referenceImages accepts up to 7 images. URLs, base64 strings, data URIs, or UUIDs from the Image Upload API. Each image is a style, a character, a location, or a storyboard beat the model preserves in the output.
• inputs.frameImages is optional, max 1, anchored to the first frame. Locks the opening composition. See Combining a first-frame anchor with a reference below.
• width and height are optional. When passed, they must both be present and match one of two combinations: 1280 × 720 (16:9 landscape) or 720 × 1280 (9:16 portrait). Audio is generated natively.
• duration is an integer from 5 to 8 seconds. Omit it and the model picks a duration that fits the prompt, which is the recommended default. You can also direct the runtime inline in the positivePrompt by naming it ("a 6-second reel", "an 8-second clip") instead of passing the parameter. It is only valid in generation mode: when inputs.video is present (the editing workflow covered in the editing video guide), duration is forbidden and the model inherits the source video's length.

Style transfer with a single reference

Pass a single reference image whose aesthetic is what you want and the model translates a brand-new scene into that style. Watercolor, claymation, ink wash, anime, oil painting, halftone print. The reference defines the look. The prompt describes the scene.

A watercolor painting of a quiet European countryside landscape: rolling green hills, a winding pale stone path, a single oak tree on the distant hillcrest. Visible paper texture beneath translucent washes of color, gentle ink line work, soft natural color bleeds

The reference shows a quiet countryside landscape. The output is a completely different scene (a farmer's market) carrying the same translucent washes, paper texture, ink line work, and pale earthy palette. The model treated the reference as a how, not a what. None of the content carried over, only the aesthetic.

Name the style ingredients in the prompt, not just the reference. Saying "the exact watercolor painting style of the reference image" plus calling out the techniques you can see ("translucent washes", "visible paper texture", "ink line work") forces the model to honor those specific signals. A vague "in the reference style" sometimes drifts toward a plausible but different aesthetic.

Character or object consistency

A single reference image of a person, mascot, or product carried across new contexts, new actions, new backgrounds. This is the workflow for branded talent, brand mascots, recurring characters in a mini-series, or product hero shots.

A female astronaut in her thirties, dark hair pulled back into a low bun, wearing a sleek white pressure suit with cyan accent stripes, a clear bubble helmet, and two circular mission patches on her left arm reading OMNI-3 and EXPLORATION

The astronaut wasn't in the reference image's pose or environment. The output puts her in an alien jungle, walking, ducking, examining a spore. The white pressure suit, the cyan accent stripes, the clear bubble helmet, and the mission patches all carry through every cut. The reference defined who. The prompt defined what.

The same workflow extends to products. A clean studio packshot in the reference position locks the bottle, the cap, the label, and the liquid colour across cuts so the model can place it into new contexts without redrawing the design every frame.

A luxury perfume bottle product shot: a tall faceted clear glass column with sharp angular edges, filled with deep amber liquid that glows warmly. A solid polished gold cap with a fine engraved geometric pattern around the rim. A small embossed gold label reads MAISON LUNA in clean serif type near the base. Centered against a soft warm cream gradient backdrop

The bottle reads as the same product on the studio backdrop, the marble vanity, and in the held hand. The faceted glass, the gold cap, and the embossed wordmark survive the cuts unchanged. This is the workflow for packshot variations of a single SKU, brand-mandated mascots that must look identical across a campaign, and any reference asset signed off by a design team.

Storyboard guidance with multiple references

When you have a specific sequence of beats you want the video to hit, pass them as multiple reference images in order. The model uses each as a visual key frame and generates the motion that connects them. This is how Google's prompt guide describes the "narrative order" use case.

The dessert plating sequence below is built from three reference frames in narrative order: the sauce being poured, the garnish being arranged, and the finished beauty shot.

A wide cinematic shot of a chef in a crisp white apron pouring a smooth stream of glossy dark chocolate sauce from a small ceramic pitcher over a vanilla panna cotta on a wide white ceramic plate

An overhead top-down shot of the chef's hands placing fresh raspberries and small edible flowers in a pattern around the dark-chocolate-glazed vanilla panna cotta, using tweezers

A close beauty shot of the finished plated dessert: the dark-chocolate-glazed vanilla panna cotta surrounded by raspberries and edible flowers, with a single small piece of edible gold leaf catching the warm overhead light

The output moves through all three beats in the order the references were passed. The pour resolves into the placement, which resolves into the beauty shot, and the motion between each pair feels natural because the references give the model a clear visual destination to interpolate toward.

State the storyboard order in the prompt explicitly. The model honors reference order more reliably when the prompt names "first reference", "second reference", "third reference" as the beats. Without that explicit mapping, three references can resolve in any order depending on the prompt content.

Combining a first-frame anchor with a reference

When the opening composition has to be exact (a product reveal, a logo splash, a hero shot signed off by a brand team), stack inputs.frameImages on top of the reference cast. The anchor locks the first frame. The references hold the cast through the rest of the clip.

A vintage cherry red convertible from the early 1960s parked on a winding coastal cliff road at golden hour. Side-on three-quarter angle showing the polished chrome bumper, round chrome headlights, and the deep red bodywork catching the warm low sun. The Mediterranean ocean stretches into the deep gold sunset horizon below the cliff

A male driver in his forties, neatly combed dark hair with a touch of grey at the temples, wearing a tan linen blazer over an open-collar white shirt, classic aviator sunglasses pushed up onto his forehead, a vintage gold wristwatch on his left wrist

The video opens exactly on the anchor: the same cherry red convertible at the same angle on the same cliff road under the same warm golden light. The driver then walks into the frame, and his salt-and-pepper hair, tan linen blazer, aviator sunglasses, and gold wristwatch all carry through from the reference. The anchor handles the opening composition. The reference handles identity.

Tips

1. Pass reference images in the order you reference them in the prompt. The model uses position as a signal. "The traveler from the first reference visits the second, third, and fourth references in order" is parsed more cleanly than the same intent left implicit.

2. Use mid-shot portraits for character references. A waist-up portrait at clean lighting and neutral background gives the model an unambiguous read on the face, wardrobe, and identifying details. Group shots, full-body, or busy backgrounds dilute the lock.

3. Name the style ingredients when doing style transfer. Don't rely on "in the reference's style" alone. Call out the technique features ("translucent washes", "visible paper texture", "ink line work") so the model honors the right signals from the image.

4. For storyboard mode, state the beat order in the prompt. Reference position alone is less reliable than reference position plus an explicit "first / cut to second / end on third" structure in the prompt.

5. Lean short and focused on the prompt itself. Google's own prompt guide calls Omni Flash "less prescriptive" than Veo. Three or four sentences naming the subject, action, and reference roles outperforms a wall of cinematic adjectives.

6. Stack a first-frame anchor when the opening is the reveal. Product launches, brand intros, logo splashes, and mini-series episodes where shot one is locked all benefit from the combined frameImages + referenceImages mode.

7. Match duration to the shot count. Omni Flash generation mode runs from 5 to 8 seconds. Single-beat shots and short transitions read cleanly at 5 to 6. Three-beat reels want the full 8. Cramming three beats into 5 seconds gives each beat under two seconds, and the cuts read as choppy.