Swap the product in a UGC clip with Gemini Omni Flash 1.1 video edit

To reuse a creator clip for a new product, send the clip as video_url to POST /v1/video-router/generate with model: gemini-omni-flash-1.1 and a prompt that describes the change, such as replacing one item with another and keeping everything else. This is the model's video_to_video edit mode, and it is meant for small changes inside a clip you already have.

What does the edit mode accept?

What does a call look like?

Native audio is always on for this model. video_url is the edit source, not a reference, so it cannot be combined with image_url, end_image_url or the reference_*_urls fields. Billing is provider list times 1.25 per output second by resolution.

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: swap-product-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "Replace the bottle in her hand with a red gift box. Keep everything else the same.",
    "video_url": "https://example.com/ugc-clip.mp4",
    "resolution": "720p",
    "mode": "async"
  }'

Why can I not add a reference image of the new product?

Because the edit source cannot be combined with image or reference inputs. The new product has to be described in words. If the product's look matters, such as a label or a specific shape, use the other route: generate a new clip with reference_image_urls of the product, or run a catalog Format that starts from your packshot.

When should I regenerate instead of edit?

Edit when the creator, the room and the movement are the part you like, and the product is a small object in the frame. Regenerate when the product is the star of the shot, because a prompt cannot show a specific label. For a fresh clip built from your packshot, the catalog Formats start from an attached image of the real product.

Either way, test on one clip before you edit a set. An edit is a paid job, so a retry with a new Idempotency-Key is charged again, while a retry with the same key and body returns the original job. A good habit is to name keys after the clip and the prompt version, so you can tell which edit you are looking at later.

What are the limits?

• Read capabilities in GET /v1/video-router/models before you build on the edit mode.
• Words describe the swap, so labels and logos on the new item are not guaranteed to match your product. Check them frame by frame.
• Audio is always generated; you cannot turn it off with generate_audio: false.
• You need rights to the source clip, and a person who appears in it still appears in the edit.
• The edit changes what is shown, so check that the result does not make a claim the original did not make.