Film noir black-and-white AI video prompt for Gemini Omni

A noir clip comes from three kinds of words: monochrome film stock, hard directional light, and a sound line with rain and one instrument. Write them into a Gemini Omni prompt, request a single continuous shot so it does not cut away, and a 10-second 720p take costs $1.25 on Sume. Add 'no dialogue' to the sound line, using the negative phrasing Google's Omni guide suggests (read 2026-10-07).

Noir hides model weaknesses well: shadows cover detail, and limited color removes a common source of error.

The look in words

Say 'black and white' once and then use lighting vocabulary rather than repeating it.

• Stock: 'high-contrast black and white, fine film grain, 1940s crime film'.
• Light: 'a single hard key light through venetian blinds, long shadows, deep blacks'.
• Weather: 'rain on a window, wet street reflecting a neon sign'.
• Camera: 'slow push-in at eye level, shallow depth of field'.
• Sound: 'rain, a distant muted saxophone, a clock ticking. No dialogue.'

Request

Keep one subject and one action. A detective at a desk lighting a cigarette is easier than a chase.

import os, requests

prompt = ('High-contrast black and white, fine film grain, 1940s crime film. '
          'A detective sits at a cluttered desk, hard key light through venetian blinds '
          'throws long shadows, rain streaks the window behind him. Slow push-in at eye '
          'level, single continuous shot. Audio: steady rain, a distant muted saxophone, '
          'a ticking clock. No dialogue.')
r = requests.post(
    'https://api.sume.com/v1/videos',
    headers={'Authorization': 'Bearer ' + os.environ['SUME_API_KEY'],
             'Idempotency-Key': 'noir-desk-1'},
    json={'model': 'gemini-omni-flash-1.1', 'prompt': prompt, 'duration': 10,
          'resolution': '720p', 'aspect_ratio': '16:9'},
    timeout=60)
print(r.status_code, r.json())

Cost per take

Sume bills list times 1.25 per second, rounded up to the cent per job; the Video Router doc lists $0.03, $0.10, $0.15 and $0.30 per second.

Common problems

If the picture comes back with color, move 'black and white' to the first words of the prompt and name the film stock. If the model puts a person in the frame who speaks, strengthen the sound line and remove any mouth or phone in the description. Native audio cannot be turned off for this model on Sume, so the saxophone and rain are decided by your words. Smoking is a safety-filter topic on some platforms; if a take is refused, replace the cigarette with a pen or a cup.

Adding a title card

Noir titles are usually white serif text on black. Do not ask Omni to render a long title inside the shot; make the card in a design tool and append it with a timeline join. Google says Omni generates readable text in videos (read 2026-10-07), which is fine for a short word on a sign in the scene, such as a neon bar name, but a title you must proofread is safer as an image. Keep the aspect ratio consistent between the clip and the card so the join needs no crop.