Gemini Omni Flash 1.1

Introduction

Gemini Omni Flash 1.1 delivers the same shot across four tiers, from a 360p draft to a 4K master, and the tier you pick is the single biggest lever on what a clip costs. A 4K second is roughly nine times a 360p second, so the tier decision belongs at the start of a workflow rather than at the end of one.

Duration is the other half of the same equation. The model produces 3 to 10 seconds per call and bills by the second, which makes tier and runtime the two numbers worth deciding before you write the prompt.

This guide covers the two ways to size a clip, when each one applies, what the tiers actually buy you, and how to pick a duration.

Two ways to size a clip

Sizing goes through either resolution or a width/height pair, and the request is rejected if you send both. They cover the same output sizes from different directions.

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: 'A slow push in on a hiker resting against a boulder on an exposed alpine ridge in the late afternoon, in a single unbroken scene. Outdoor apparel brand cinematography. Low warm sun raking across lichen-covered rock. The audio is steady wind across the ridge, no music.',
  resolution: '1080p',
  duration: 6
})

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": "A slow push in on a hiker resting against a boulder on an exposed alpine ridge in the late afternoon, in a single unbroken scene. Outdoor apparel brand cinematography. Low warm sun raking across lichen-covered rock. The audio is steady wind across the ridge, no music.",
            "resolution": "1080p",
            "duration": 6
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "c8d3e2f5-9a0b-4134-c2d3-e4f506172839",
      "model": "google:gemini@omni-flash-1.1",
      "positivePrompt": "A slow push in on a hiker resting against a boulder on an exposed alpine ridge in the late afternoon, in a single unbroken scene. Outdoor apparel brand cinematography. Low warm sun raking across lichen-covered rock. The audio is steady wind across the ridge, no music.",
      "resolution": "1080p",
      "duration": 6
    }
  ]'

runware run google:gemini@omni-flash-1.1 \
  positivePrompt="A slow push in on a hiker resting against a boulder on an exposed alpine ridge in the late afternoon, in a single unbroken scene. Outdoor apparel brand cinematography. Low warm sun raking across lichen-covered rock. The audio is steady wind across the ridge, no music." \
  resolution=1080p \
  duration=6

{
  "taskType": "videoInference",
  "taskUUID": "c8d3e2f5-9a0b-4134-c2d3-e4f506172839",
  "model": "google:gemini@omni-flash-1.1",
  "positivePrompt": "A slow push in on a hiker resting against a boulder on an exposed alpine ridge in the late afternoon, in a single unbroken scene. Outdoor apparel brand cinematography. Low warm sun raking across lichen-covered rock. The audio is steady wind across the ridge, no music.",
  "resolution": "1080p",
  "duration": 6
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "c8d3e2f5-9a0b-4134-c2d3-e4f506172839",
    "videoUUID": "5b6c7d8e-9f0a-4123-b4c5-d6e7f8091234",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/5b6c7d8e-9f0a-4123-b4c5-d6e7f8091234.mp4"
  }
]

resolution takes 360p, 720p, 1080p, or 4K and names a tier without committing to an orientation. When the call also carries input media, the tier follows the aspect ratio of that media, which is what makes it the right choice for editing and extension work.

width and height travel together, both or neither, and must match one of eight supported pairs:

Landscape and portrait are the only orientations. There is no square output and no arbitrary pair, so a value off this table fails validation rather than snapping to the nearest legal size.

Portrait is a first-class output rather than a crop of a landscape render, which matters for social work where the subject has to sit in a tall frame from the start.

What each tier buys

The four tiers below ran the same prompt, one call each. Because the model exposes no seed, these are four separate takes rather than one take rendered four ways, so read them for fidelity rather than for a frame-by-frame match.

Web delivery flattens most of that difference, because every clip on this page is re-encoded for the browser. The pair below is the honest version: one frame pulled from the 360p take and one from the 4K take, cropped to the same region and shown at their own pixel density.

The fabric weave and the detail in her face survive the crop at 4K and dissolve into flat colour at 360p. That gap is the whole argument for the tier ladder: 360p is for judging a take, not for judging a texture.

Drafting cheap, finishing expensive

Per second of output, the tiers price at $0.036 for 360p, $0.10 for 720p, $0.16 for 1080p, and $0.32 for 4K. A six-second clip therefore runs about $0.22 at the draft tier against $1.92 at the master tier.

That spread is what makes an iterate-then-commit loop worth building. Run prompt variations at 360p until the framing, the action, and the audio are right, then re-run the winning prompt once at the delivery tier. Ten drafts plus one 4K master costs less than three 4K attempts and gets you more looks at the idea.

Choosing a duration

duration accepts a whole number of seconds from 3 to 10 and defaults to 6. It is not a quality control, it is a container: the model paces whatever you asked for to fill the time you gave it.

The three-second clip carries one action and ends on it, which is the right shape for a product-detail loop or a UI transition. The ten-second clip carries a sequence with a beginning and an end, and the florist prompt names four steps that the model has room to pace out.

Ask for four actions in three seconds and each one gets under a second, which reads as rushed. Ask for one action in ten seconds and the model pads, usually with camera drift. Match the number of beats in the prompt to the seconds you bought.

To go past ten seconds, generate a clip and extend it. See the extension guide.

Tips

1. Pick resolution or width/height, never both. Sending both fails validation. The preset is simpler when orientation is already implied, the pair is explicit when it is not.

2. Use resolution for anything with input media. It takes its aspect ratio from the source, and width/height are rejected alongside inputs.video anyway.

3. Draft at 360p. At roughly a ninth the per-second cost of 4K, the draft tier is where prompt wording gets settled. Commit to a delivery tier once.

4. Don't judge texture from a draft. Fabric, skin, foliage, and fine type all collapse at 360p. Judge framing, action, and audio there, and judge detail at 1080p or above.

5. Generate portrait natively. The 9:16 pairs are real output sizes, and framing a subject for a tall frame beats cropping a landscape render later.

6. Match beats to seconds. One action fits 3 seconds, a short sequence fits 10. Overstuffing a short clip rushes it and understuffing a long one invites drift.

7. Remember the 10-second ceiling. It is the maximum for a single call, and longer pieces come from extending a clip rather than from a bigger duration.