Kitchen gadget UGC demo: 15 seconds on Seedance 2.5 with references

For a gadget demo, send the product photo as an image reference, describe one action in the prompt, and keep the clip to 15 seconds. On Seedance 2.5, a 15-second 9:16 clip at 720p costs $8.67 on Sume, and a 480p proof costs $4.04.

A demo has one job: show the product doing the thing. One gadget, one action, one result. A garlic press crushes a clove; a salad spinner spins. Anything beyond that belongs in a second clip.

What to send as references

Sume's docs say a model accepts a reference type only if its supported_input_references lists it, and the Seedance 2.x models accept image, video and audio references. The request validation caps Seedance 2.5 at 9 reference images, 3 reference videos and 3 reference audio files, with 12 references in total. A gadget demo needs far fewer.

Send one to three product photos: front, in use and a close-up of the detail. Add a short audio clip only if you have a voice you are allowed to use. ByteDance Seed's launch post lists up to 30 images, 10 videos and 10 audio clips per pass for Seedance 2.5 (read 2026-10-05); those are vendor numbers, not Sume's.

The demo prompt

Use the references by role, not by file name. Say what each one is for.

• Setting: "A home kitchen counter, morning light, a person in a gray sweater, only their hands and torso visible."
• Product: "The stainless steel garlic press from the reference photo."
• Action: "They place one garlic clove in the press, squeeze the handles once, and a small amount of crushed garlic comes out onto a cutting board."
• Camera: "Handheld phone framing from slightly above, steady, no zoom."
• Sound: "Kitchen room tone, a soft squeak of the handles. No music."

Price at 15 seconds

Same length, three resolutions, 9:16.

Request with image references

The input_references array is the reference-to-video field in the Sume video docs.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: garlic-press-001" \
  -d '{"model":"seedance-2.5","prompt":"A home kitchen counter, morning light. A person in a gray sweater, hands and torso only, places one garlic clove in the stainless steel press from the reference photo and squeezes once. Handheld phone framing from slightly above. Kitchen room tone.","input_references":[{"type":"image_url","image_url":{"url":"https://example.com/press-front.jpg"}}],"duration":15,"resolution":"480p","aspect_ratio":"9:16"}'

Shot list for a 15-second demo

Plan the clip in three beats so the model has a clear order. Beat one, 0-4 seconds: the product sits on the counter and a hand reaches for it. Beat two, 4-11 seconds: the single action, done once, at a natural speed. Beat three, 11-15 seconds: the result in frame, held still so a viewer can read it. If you write the beats into the prompt as plain sentences in that order, the clip tends to follow them; if you write a paragraph of adjectives, the order is left to chance.

Keep the person generic. A UGC look comes from the framing, the room tone and the imperfect handheld motion, not from a named creator. If you want a specific on-screen person, that is a different job with its own rights questions and should not be done by describing someone's face in a prompt.

What to verify

Look at the product in the clip against your photo: handle shape, color, number of parts. A reference guides the look but does not pin a frame, so small differences are normal, and they matter for a product that buyers will compare with the clip. If you need the product exact from the first frame, use a first frame instead; the docs say that when both fields are sent, frame_images controls the mode.

Also check what the hands do. Hands holding a small object are a common place for odd results. If a take fails there, simplify the action to one hand, shorten to 10 seconds ($2.69 at 480p) and retry. Cheap drafts exist for this reason.