How do I put three products in one AI video from reference photos?

To put three products in one AI video, send the three photos as input_references of type image_url to gemini-omni-flash-1.1 on POST /v1/videos, and describe the scene with all three. On Sume, an 8-second bundle clip at 720p is $1.00. Three separate 5-second 720p clips would cost $1.89 and would still need to be joined.

A bundle clip is a good fit for gift sets, a table of items or a before-and-after of a kit. It is a poor fit when each product needs its own full screen time. Then use one clip per product and a render.

What input_references does

Reference images guide the content. They are not fixed frames. That is the key difference from frame_images, which pin the first or last frame. With references the model decides where each product appears, so you describe the scene and tell it which reference is which.

The Video Router docs give the addressing used for Omni: <IMAGE_REF_0>, <IMAGE_REF_1> and so on, with up to 10 images and 3 clips of at most 3 seconds each. Use them in the prompt in the order you send the images. Omni takes image and video references, not audio. Because you cannot pin a frame in a reference job, plan on a retake or two.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: bundle-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "On a wooden table, <IMAGE_REF_0>, <IMAGE_REF_1> and <IMAGE_REF_2> sit side by side, soft window light, slow push-in",
    "duration": 8,
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://media.sume.com/artifacts/artf_demo/a.jpg" } },
      { "type": "image_url", "image_url": { "url": "https://media.sume.com/artifacts/artf_demo/b.jpg" } },
      { "type": "image_url", "image_url": { "url": "https://media.sume.com/artifacts/artf_demo/c.jpg" } }
    ]
  }'

Bundle or three singles

One 8-second bundle at 720p is $1.00. Three singles at 5 seconds are $0.63 each, $1.89 in all, and they need a Timeline render at $0.10 to become one video. The bundle is cheaper, but each product gets less detail. Check each product's colour and shape in the result: if a product has changed, go back to singles with a first-frame image for each.

A middle path: one bundle clip for the opening, then three short singles for the detail shots. That is four jobs and a render.

Write the prompt for references

Say where each product goes and what it does. 'On a wooden table, <IMAGE_REF_0> on the left, <IMAGE_REF_1> in the centre and <IMAGE_REF_2> on the right, soft window light, slow push-in' gives the model a layout to follow. Without a layout it may crowd the three products or leave one out of frame.

Keep brand names and printed text out of the prompt. If a label has to be readable, check it in the result at full size; if it is not, overlay a still of the real label, as the label-fidelity post on Sume's blog explains, or burn the product names in with caption cues.

Keep the camera move simple. A slow push-in or a small orbit across eight seconds shows three products without losing the layout, and a fast move is the quickest way to lose one of them.

What to check in the result

Reference images guide a clip, they do not pin it. Look at the finished video for each product in turn: is the shape right, is the colour right, is any text on the pack still readable? A model can blend two products or drop one. If one is missing, say so in the prompt by name and run it again, since a retake is just another clip at the same price.

Do not rely on a model to draw labels or small print. If a label must be exact, overlay your own still with Timeline compose, as in the stored note on keeping a label exact, and check the final file at full size on a phone.

Keep the three reference photos and the prompt with the campaign. If you add a fourth product next month, you start from the same set of pictures and one new photo.

Cost by resolution

• Omni accepts 16:9 and 9:16 only, so a square listing needs a crop in the render.
• Draft at 360p, then re-run only the keeper at the final size.
• Send the photos on a plain, even background so each product reads as one object.