Image-to-video prompts: separate action, camera movement and sound

Choose the input mode, then describe one continuous shot

Start with the first-frame, first/last-frame or smart-reference method supported by the current model. Describe subject action, camera movement and the ending state separately. Make the first attempt a single continuous shot so failures are easier to identify.

A camera orbiting a product differs from a product rotating in front of a fixed camera. State what moves, in which direction and how far. A large orbit requires the model to invent surfaces absent from the reference. If those surfaces must be accurate, prepare more usable references or reduce the viewpoint change.

First/last frames and smart references serve different roles

• First/last-frame mode uses dedicated image slots for the starting and ending states. Whether you can use just a first frame or also an ending frame depends on the workflow. The ending frame is not a general style reference.
• Smart-reference mode accepts images, video or audio according to model support, to guide subjects, motion or sound. Adding two images does not set a first and last frame.
• After changing mode or model, recheck input types, counts, duration, aspect ratio, resolution and audio parameters. These input and output capabilities are not available on every model.
• Image imports still have a 20,250,000-pixel total limit and an 8192px longest-edge limit, with a resize-or-cancel choice when exceeded. Check the current parameter guidance for video and audio limits.

Organize the shot with five details

• Subject: who or what is in the shot, and which appearance details must stay.
• Action: its sequence, direction, range and ending pose.
• Camera: fixed, push-in, lateral movement or orbit. Choose one main movement for the first attempt.
• Environment: location, light direction and background changes that are allowed.
• Sound: speakers, dialogue, ambience and music requirements when the model supports them.

Go from reference to finished clip

1. Choose video generation at the top left of Composer, then select an input mode and an available model.
2. Add uploaded references and check the first/last-frame slots or smart-reference items and their order.
3. Describe one continuous action and state the identity, product shape and lighting to preserve.
4. Set supported duration, aspect ratio, resolution and audio options through the controls. Check the price, then submit.
5. Play the complete result and note failure timestamps. Change only the input, action or camera instruction in each revision.

Author-written single-shot prompt, not generation-tested

This example has not been tested in a generation run. Include sound instructions only if the model supports audio generation; otherwise retain them as notes for later sound production. A prompt cannot enable duration, resolution or audio capabilities absent from the controls.

Preserve the reference person, clothing and room. The person looks at the camera on the table, gently picks it up, and ends holding it at chest level. Fixed camera, medium framing, one continuous shot. Stable window light. Only subtle clothing and room sounds, with no dialogue or music.

Common problems and how to revise

• Subject and camera both move unpredictably: specify a fixed camera or keep one camera movement, and reduce simultaneous actions.
• The transition between endpoints is abrupt: check that subject, position and lighting can connect plausibly, then simplify the action. An ending-frame constraint does not guarantee every intermediate frame.
• Hands intersect objects or product structure changes: note the timestamp, reduce movement or occlusion, and check the original reference.
• No sound: first check model support and the audio setting, then distinguish a generation issue from a muted player.

Review the complete timeline

Check whether the opening preserves the reference subject, whether the middle has intersections, deformation or light jumps, and whether the ending works for the next shot. For results with sound, use headphones to check dialogue, ambience and synchronization. A poster or best frame is insufficient.

Download the final video, check its actual dimensions, duration and sound, then preview the compressed result on the destination platform. Save references, prompts and the finished clip separately; browser history is not permanent media backup.