Add Halloween decorations to a shop video with an Omni edit prompt

Send the finished clip as video_url to gemini-omni-flash-1.1 on the Video Router, and write a prompt that names the decorations to add and what must stay the same. You do not re-shoot, and you do not need a reference image. Sume's edit mode takes the clip plus a text instruction and nothing else, so the prompt carries the whole brief. That also means the quality of the result depends on how precisely you describe placement, which is why the template below spends most of its words on where things go.

Halloween is under four weeks away, which makes this the week to test it. A shop that filmed a September window or counter clip can turn it into a seasonal version for the cost of one edit, instead of a second shoot.

The request

Sume documents the edit shape as video_url plus a prompt, with resolution optional (the default is 720p). Do not send aspect_ratio or duration, and do not combine video_url with image_url, end_image_url or any reference_*_urls. The Sume docs also say the API rejects generate_audio: false, because native audio is always on for this model.

A prompt that follows the docs' own pattern ("Replace the bottle with an apple. Keep everything else the same.") works as a template: state the change, then state the constraint.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: halloween-window-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "Add paper bats on the window glass, a small row of carved pumpkins on the counter and orange string lights along the top edge. Keep the people, the camera movement, the products and the lighting exactly as they are. Do not add any text.",
    "video_url": "https://example.com/window-clip.mp4",
    "resolution": "1080p",
    "mode": "async"
  }'

What it costs

Edit output is billed per output second at the resolution you ask for, and the output length follows the source clip. The duration field is only a reserve hint for edits.

Prompt choices that matter

Name each prop once and place it. "Pumpkins on the counter, left of the register" gives the model a spot; "make it spooky" gives it permission to change the whole mood, including the lighting you wanted to keep.

• List three to five props, not ten. Every extra item is another place where the edit can drift from the source.
• Say "do not add any text" if the clip includes signage you will re-letter yourself.
• Keep the constraint sentence (people, camera, products, lighting) in every variant so a batch of locations stays comparable.
• Expect the edited clip to carry audio. If the original had a music bed you want to keep, compare the audio before you ship.

Check before you publish

Pull stills from the original and the edit at the same timestamps and put them side by side. The video frames route takes explicit at times, so you can compare 0 s, the middle and the last second. Look at hands, product labels and faces first, because those are where an edit that was told to keep everything usually slips.

Related how-tos: adding fireworks or confetti with the same edit mode and changing the weather in a clip. If your source is much longer than the model's 3 to 10 second generation range, trim it first.