Tasting menu opener: 40 s, four courses, Omni first/last frames

A 40-second tasting-menu opener on Sume is four Gemini Omni 1.1 Flash clips, one per course, each with an empty-plate photo as image_url and the plated-dish photo as end_image_url. The model fills in the motion between the two frames. Draft the four at 360p, approve, then render 1080p finals and join them with Timeline 1.0. The worked example below prices it at about $1.50 for drafts, $7.50 for finals and $0.10 for the timeline.

Why first and last frames fit a dish

A plated dish is a known end state. The chef already knows what the plate looks like when it leaves the pass. First/last frame control lets you fix both ends of a shot and leave the in-between (hands, sauce, steam) to the model (Google, read 2026-10-05). On Sume the Omni docs list image_url with an optional end_image_url as one request shape. end_image_url needs an image_url: you cannot send a last frame alone (Sume docs: Video Router, read 2026-10-05).

The photos you need

Matching angle and light between the pair is what keeps the clip believable. If the plate moves in the photos, the model has to invent the camera move too, and a draft will show whether it did so well.

• Four empty plates or bowls, shot from the same angle as the dish.
• Four plated dishes from that same angle, same light, same table.
• Public HTTPS URLs for all eight images.

Cost

The estimator's provider list rates ($0.03 per second at 360p, $0.15 at 1080p) times Sume's 1.25 are $0.0375 and $0.1875 per second. Check the live rates on the pricing page and in the catalog.

Assemble

Slots start at 0, 10, 20 and 30 with 10-second durations; the 40-second audio spine is a track you choose, or audio.mode: silence if you add sound later. Use fit: cover for a 16:9 source in a 16:9 output; the timeline default output is 1080x1920, so set output.width and output.height for landscape (Sume docs: Timeline 1.0, read 2026-10-05).

A fade of 0.25 seconds between courses reads as a deliberate course change. Put fades only after the first slot. Check the draft of each course for the one thing a menu cannot afford: a dish that changes. If the draft's plated end frame differs from your photo, re-prompt or re-run that course before you pay for 1080p.

When the pair does not fit

A dish that is assembled at the table, such as a sauce poured from a jug, has an interesting middle and a simple end. A dish whose empty plate and finished plate look nearly the same gives the model little to do, and the clip may feel static. In that case drop end_image_url and use a single first frame with a prompt that describes the pour.

Check every draft against the plated photo, not against your memory of the dish. Garnish placement, sauce color and plate rim are the places where a generated end frame drifts. Because the model interpolates between your two frames, a draft that lands near the photo is a good sign, and a draft that lands far from it is a reason to change the prompt before you pay for 1080p.