Prompts and reference images
A useful prompt describes a shot you can recognize when you see it. Start with the action and composition that matter to your project, then add only the details needed to direct them. More words are not automatically more control.
Describe the shot
Use a simple order: subject, action, setting, camera, and light. You do not need to include every item in every prompt, but this order helps you notice what is missing.
- Subject: Name the main person, object, or place. Keep descriptions specific enough to distinguish it from the background.
- Action: Describe one main movement. “The cyclist turns toward the sea” is easier to assess than “a dramatic moment of freedom.”
- Setting: Add the environment that affects the shot, such as an empty coastal road or a crowded station.
- Camera: Choose a clear position and movement. A wide static shot and a close tracking shot serve different purposes.
- Light: Use a concrete cue such as soft window light, overcast daylight, or low evening sun.
For example:
A cyclist pauses beside a quiet coastal road at sunrise. A wide, static shot shows the ocean behind them. Their jacket moves gently in the wind. The cyclist remains still and looks toward the water.
The camera stays still while the jacket moves. That distinction makes the intended motion clear. If you need a closer view, change the camera instruction rather than adding “close-up” after “wide shot.” The Creator prompt field currently allows up to 500 characters, so spend that space on visible details and clear actions.
Use a reference image
In Creator, references are connected through saved elements, such as a character, location, or prop. When the element picker is available, type @ in the prompt or use the @ button to select an element. Choose the saved item from the picker; typing a name as ordinary text does not attach its images.
If you create a new element, add reference images during its setup before using it. The interface supports adding uploaded or generated images to an element. Check the images and description together: they should identify the same subject, rather than ask the model to reconcile unrelated appearances.
- For a character, choose clear views of the appearance you want to carry into the scene. Avoid a set of images with conflicting outfits or markedly different faces unless that variation is intentional.
- For a prop, make its shape and important details easy to see. A busy background can distract from the object you are trying to preserve.
- For a location, choose views that communicate the layout and atmosphere relevant to your shot. Add motion instructions in the prompt instead of assuming the image explains how the camera should move.
Reference support and image limits depend on the selected model. Follow any format, count, or registration message shown before generation. If an element’s image set changes, a model may require its registration to be updated. The @ picker being available does not mean every model accepts every reference configuration.
A reference guides the result; it does not guarantee an exact identity, composition, or object shape in every frame. If the appearance is critical, inspect the entire clip rather than relying on its first frame.
Avoid conflicting requests
Keep the shot achievable within its chosen duration. A prompt that asks someone to enter a room, pick up a cup, cross to a window, and leave may need several shots. Split the sequence into separate generations, then arrange the clips in your editing project.
Replace vague quality words with the correction you need. “Make it more cinematic” could change many things; “keep the camera at table height and use soft side light” gives you a clearer result to judge. Likewise, asking for a still camera and a fast orbit in the same prompt creates competing directions.
For the cup example in your first video, a focused revision might be:
The camera stays still. Only the steam moves slowly upward. The cup remains fully visible throughout the shot.
Use that revision to replace the earlier motion instruction, not to append an opposing instruction to it.
Learn from each attempt
Before generating again, name the issue in one sentence: “The subject leaves the frame,” or “The motion starts too abruptly.” Change the relevant part of the prompt and keep a note of the result. Keep a successful version before exploring a different direction.
If the same problem persists, simplify the action, check whether the reference conflicts with the description, and review whether the model suits the task. Set a limit on further attempts before continuing. See costs and usage for planning iterations and privacy and sharing before using portraits or client material.