Gemini Omni Flash 1.1

Introduction

A prompt describes a person. A reference is the person. Gemini Omni Flash 1.1 takes up to 7 reference images and 3 reference videos, and treats them as ground truth for who is in the shot, what the product looks like, or which visual language the frame is drawn in.

That distinction is what makes the mode worth reaching for in commercial work. Brand talent, a signed-off packshot, a mascot, and an approved illustration style all have to come back identical across a campaign, and a prompt alone re-rolls them every call.

Three references, one call: the trainer, her bottle, and the room she trains in, all held to what the stills say. This guide covers the request shape, the three jobs a reference image does, what a reference video adds, and how the two combine.

The request

References go in inputs, and the prompt says what each one is for.

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-1.1',
  positivePrompt: 'The trainer from the first reference image runs a session in the studio from the third reference image. She finishes a set of kettlebell swings, then picks up the sage-green bottle from the second reference image. Keep her appearance, the bottle design, and the studio identical to the references.',
  inputs: {
    referenceImages: [
      'https://example.com/trainer.jpg',
      'https://example.com/bottle.jpg',
      'https://example.com/studio.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-1.1",
            "positivePrompt": "The trainer from the first reference image runs a session in the studio from the third reference image. She finishes a set of kettlebell swings, then picks up the sage-green bottle from the second reference image. Keep her appearance, the bottle design, and the studio identical to the references.",
            "inputs": {
                "referenceImages": [
                    "https://example.com/trainer.jpg",
                    "https://example.com/bottle.jpg",
                    "https://example.com/studio.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": "f1a6b528-2d3e-4467-f506-172839405162",
      "model": "google:gemini@omni-flash-1.1",
      "positivePrompt": "The trainer from the first reference image runs a session in the studio from the third reference image. She finishes a set of kettlebell swings, then picks up the sage-green bottle from the second reference image. Keep her appearance, the bottle design, and the studio identical to the references.",
      "inputs": {
        "referenceImages": [
          "https://example.com/trainer.jpg",
          "https://example.com/bottle.jpg",
          "https://example.com/studio.jpg"
        ]
      },
      "width": 1280,
      "height": 720,
      "duration": 8
    }
  ]'

runware run google:gemini@omni-flash-1.1 \
  positivePrompt="The trainer from the first reference image runs a session in the studio from the third reference image. She finishes a set of kettlebell swings, then picks up the sage-green bottle from the second reference image. Keep her appearance, the bottle design, and the studio identical to the references." \
  inputs.referenceImages.0=https://example.com/trainer.jpg \
  inputs.referenceImages.1=https://example.com/bottle.jpg \
  inputs.referenceImages.2=https://example.com/studio.jpg \
  width=1280 \
  height=720 \
  duration=8

{
  "taskType": "videoInference",
  "taskUUID": "f1a6b528-2d3e-4467-f506-172839405162",
  "model": "google:gemini@omni-flash-1.1",
  "positivePrompt": "The trainer from the first reference image runs a session in the studio from the third reference image. She finishes a set of kettlebell swings, then picks up the sage-green bottle from the second reference image. Keep her appearance, the bottle design, and the studio identical to the references.",
  "inputs": {
    "referenceImages": [
      "https://example.com/trainer.jpg",
      "https://example.com/bottle.jpg",
      "https://example.com/studio.jpg"
    ]
  },
  "width": 1280,
  "height": 720,
  "duration": 8
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "f1a6b528-2d3e-4467-f506-172839405162",
    "videoUUID": "8e9f0a1b-2c3d-4456-e7f8-091234567890",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/8e9f0a1b-2c3d-4456-e7f8-091234567890.mp4"
  }
]

• inputs.referenceImages accepts up to 7 images as URLs, UUIDs, data URIs, or base64. Each is capped at 5 MB and 4096 × 4096.
• inputs.referenceVideos accepts up to 3 clips as URLs or UUIDs, each up to 30 seconds.
• The two can be sent together, which makes them the only pair of media inputs this model allows in one call.
• Sizing and duration work exactly as they do for a plain text-to-video call.

What a reference image does

The same parameter covers three jobs, and the prompt is what tells them apart.

What a reference video adds

A reference video carries a character or an object, the same way an image does, with one difference that matters: it also shows the model how the thing moves. For a mascot or a product with moving parts, that motion signature is most of the identity.

The clip below is the source. It exists only to define the mascot.

Passed as inputs.referenceVideos, it puts the same mascot somewhere new, with the waddle and the antenna bob intact.

Combining images and videos

Reference images and reference videos are the only two media inputs this model accepts together, and the pairing has an obvious use: a referenced character handling a referenced product.

Say which reference is which when a call mixes both. "The mascot from the reference video presents the headphones from the reference image" leaves nothing to infer, and that phrasing is what keeps the model from blending two identities into one object.

Writing the prompt

Three habits carry most of the quality in this mode.

Point at references by position. "The trainer from the first reference image", "the studio from the third reference image". Order is a signal the model reads, and naming it removes the ambiguity of three stills arriving at once.

Close with what must not change. A short pin clause listing the wardrobe, the finish, the label, or the palette is what holds a design through a scene the model is otherwise free to invent.

Frame a portrait as a mid-shot. Waist-up framing on a neutral background with the face and wardrobe clearly visible gives the cleanest identity lock. Full-body shots, group shots, and busy backgrounds all dilute it.

Tips

1. Up to 7 images and 3 videos, and they can travel together. No other pair of media inputs on this model can.

2. Reference by position in the prompt. "The first reference image" beats leaving the model to guess which still is the character and which is the location.

3. Use mid-shot portraits for people. Waist-up, neutral background, face and wardrobe visible. That framing locks identity better than anything else you can control.

4. For style, name the ingredients. Palette, outline weight, flat fills, grain. Pointing at the reference alone drifts toward a plausible but different look.

5. Reach for a reference video when motion is part of the identity. A mascot's walk or a product's articulation is carried by the clip in a way a still cannot express.

6. Expect no audio from a reference video. The track is regenerated from your prompt every time, so describe the sound you want.

7. Close every prompt with a pin clause. Name the wardrobe, the finish, the label, or the palette that has to survive the new scene.

8. Don't mix references with frames or a source video. Those modes are exclusive. Run the reference pass first, then pin frames or edit in a second call.