First-last frame is the product-turn model

When the job is object rotation or a move from A to B, first-last frame is the control. Text-to-video is not. Do not prompt a turntable as a paragraph.

Flux 3 First Last Frame to Video interpolates between a still you own and a still you own. You pin the starting pack shot and the ending pack shot. The model solves only the motion. That is why a bottle can turn, a lid can seat, a shoe can spin, a device can go from closed to open, without the label becoming a cousin of itself.

The loops-and-transitions review is a different post: Flux 3 first/last frame. This page is the product-turn router. If you came here for a seamless steam loop, go there. If you came here because the hero SKU has to rotate and still be the SKU, stay.

Why a paragraph cannot turn a product

Text-to-video receives a description of a bottle and invents a bottle. Sometimes it is close. Close is how you get an ad you cannot ship: the cap is a millimetre off, the lockup is re-typeset, a seam appears where the die-cut has none. Re-rolling the paragraph does not add a contract. It adds more invented bottles.

Image-to-video from one still is better — you keep the start — and then the model guesses the end. On a turn, the guess is the risk. Halfway through the orbit the label drifts, and you cannot pull it back because you never named the landing frame.

First-last names both ends. Front of pack as frame one. Three-quarter or reverse as frame two. The orbit is now a bounded problem: travel from this photograph to that photograph in N seconds. The product video generator is built around that contract. The photographs are the product. The prompt is the camera.

How to brief a turn

1. Shoot or generate two honest stills. Same object, same lighting family, same lens height. Front and three-quarter. Closed and open. On-stand and in-hand. The stills have to be connectable in the duration you will ask for. A macro of the cap to a wide of the aisle is not a turn, it is a warp.
2. Do not describe the product in the prompt. The stills already did. Describe the move: slow table-top orbit, locked tripod, no handheld, no morph, product remains the same object.
3. Keep the product large in frame. Fine print is the first thing a generator loses. If the label must read, it has to fill the frame at both ends.
4. Inspect the midpoint. First-last holds the ends by construction. The middle is where a turntable cheats — extra fingers on a cap, a label that swims. If the middle is wrong, reroll the motion, not the stills, unless a still was the lie.
5. One axis per generate. A 180° yaw or a lid seating. Two axes at once is how the model blends a turn with a transform.

Duration on Flux 3 first-last runs 5 to 20 seconds at 720p or 1080p, with native audio. For a product turn you usually want the short end and "no music" in the prompt, because you will score the plate later. A 20-second orbit of a bottle is a different film than a 6-second pack turn for a PDP loop.

What is not a product turn

A to B as a scene change. Kitchen still → store still is a transition, not a turn. That can still be first-last, and it is the other post. Here the object has to survive as itself.

A talking presenter holding the SKU. The mouth is the shot. Generate the presenter elsewhere and cut. First-last will try to interpolate a face and a pack at once and will do neither well.

A paragraph about "cinematic product video, slow orbit, rim light." That is text-to-video, and it will invent the SKU. If you do not have two stills, you do not have a first-last job yet. Make the stills. Then spend the video credit.

A turn you will recrop into five ratios after the fact. Generate the ratio you will ship, or generate vertical first. First-last inherits the composition of the stills. A 16:9 pack shot squeezed into 9:16 loses the label you pinned.

The decision rule

Ask: do I know the start picture and the end picture, and is the point to travel between them without inventing a new object?

Yes → first-last. Pin both stills. Prompt the camera only.

I know the start only, and the end can wander → image-to-video, and accept that the pack may drift.

I do not have a picture of the object → you are not in product-turn yet. Photograph it, or build stills you will actually stand behind, then come back.

I want more of a take I already like → that is extend, not first-last. Extend has no destination frame. A turn is a destination.

The expensive mistake is treating the turn as a prompt-engineering problem. It is an input-shape problem. Two photographs beat a better paragraph. The Flux 3 first-last endpoint is the endpoint that accepts that shape. Use it for the orbit. Use something else for the world around the orbit, and cut.

FAQ

Can I prompt "slow turntable of the bottle" in text-to-video and skip the stills?

You can. You will get a bottle. It will not be the bottle often enough to ship. First-last exists because the stills are the spec. A paragraph is not a spec for a SKU.

Do both frames have to be photographs of the real product?

They have to be the object you are willing to publish. Real photos are the honest path. Generated stills work if you have already locked them as the pack — same as locking a hero still before any video credit. If the stills disagree about the label, the turn will hallucinate a compromise.

Is this the same as a seamless loop?

A loop is first-last with the same frame in both slots. A product turn is first-last with two different frames of the same object. Same endpoint, different job. Use the loop recipe when the PDP needs idle motion. Use the turn recipe when the buyer needs to see the other side.

What if the product has to speak or the scene has to be a kitchen?

Split the file. First-last for the pack orbit. A talking or scene model for the rest. Comp them. Asking one interpolate to be a turn and a presenter and a room is how the SKU stops being the SKU.