Luma AI image to video can animate one starting image or connect a start image to an end keyframe. Choose the model before planning the duration: Luma’s Ray3.14 documentation lists five-second image-to-video generation, while its older Ray2 guide lists five- or ten-second choices. Generate and inspect the first shot before using Extend to continue it.
If the subject is an original companion character, Cherrypop’s creator lets you define its appearance and personality before moving into media. The Luma workflow here focuses on shot endpoints and continuation; neither route guarantees visual continuity.
Begin with the destination of the shot
Consider a photograph of a small greenhouse with its door closed. You want a short clip that ends with the door open and the interior visible. The end condition matters because it determines whether the clip can connect to the next scene.
There are two ways to plan the task. You can start from one image and describe the opening action, or prepare an ending image that establishes the desired final composition. The second method requires more source preparation but makes the destination explicit.
Luma’s keyframes guide describes uploaded or generated start and end images, optional text direction, and extending toward an image keyframe. The practical implication is that source preparation and motion direction are separate jobs. An end frame should resolve the destination, not introduce an unexplained redesign of the scene.
| Planning question | Greenhouse example | What to prepare |
|---|---|---|
| What opens the shot? | Door closed, greenhouse fully visible | A clear starting composition |
| What must change? | Door opens enough to reveal the interior | One action with a readable result |
| What should remain recognizable? | Frame, glass panels and path | Compatible details in both images |
| Where does the next shot begin? | A view toward the entrance | An ending suitable for an edit |
Inspect the generated movement against the door’s structure, particularly its hinges and opening direction.
Choose your Luma AI image to video model
Luma’s Ray3.14 user guide distinguishes image-to-video from text-to-video duration. The video models field guide also confirms start and end keyframes for Ray3.14. Use the row for your selected model rather than combining settings across guides.
| Documented model | Frames and duration | Extension and finishing |
|---|---|---|
| Ray3.14 | Start/end keyframes; five-second image-to-video. Five or ten seconds applies to text-to-video. | Extension up to about 30 seconds. Audio is a separate step; Reframe is unsupported in its guide. |
| Ray2 | Start/end keyframes; five- or ten-second choices in its June 2025 FAQ. | Extend is documented, with a warning about quality loss in longer chains. |
The Ray2 FAQ describes paid-subscriber access. For Ray3.14, the user guide says it is available to subscribers, while the Ray3.14 FAQ additionally allows Free accounts to use Draft Mode. That is a limited documented access route, not a free allowance for every resolution or operation. Confirm the mode and charge in your account.
Two older instructions need care. The November 2024 keyframe tutorial starts from four seconds; that is not a current all-model rule. Ray2’s FAQ calls 30 seconds a cap but also discusses quality dropping after that point and examples of longer generations. Its wording does not establish a dependable longer maximum. Check the active model’s limit before planning a long continuous shot.
A keyframe is also different from a character reference. Ray3.14’s guide lists keyframes as supported but excludes its reference feature. An image defining the first or last frame is not evidence that a separate identity-reference control is available.
Prepare a plausible transition
For the greenhouse, compare the first and last images before entering a prompt. Does the door hinge appear on the same side? Is the path in the same place? Has the camera moved to an incompatible viewpoint? Did the interior change simply because the source images came from different versions of the design?
Resolve accidental differences first. Then describe the intended action in plain language. An original instruction is: “The greenhouse door opens slowly toward the path, revealing the interior. The camera stays outside at the same height.”
Luma’s field guide recommends describing what changes between keyframes instead of repeating static details. For this example, use the text to explain the door movement and camera position; the images already establish the greenhouse’s appearance.
If the desired ending exposes a large area missing from the starting image, decide whether you are comfortable with inferred details. For a fictional scene, invention may be acceptable. For an accurate representation of a real location, another production method or better source coverage may be more appropriate.
Build the shot and extend toward an image
Use your own source images or assets you have permission to animate. Luma’s keyframe and Extend instructions describe the following operation sequence. The button labels come from those tutorials; their older duration examples should not override the model-specific limits above.
- Select your model and supply the opening image. Add an ending image if the final composition matters. Luma documents uploading images or selecting ones generated in Dream Machine.
- Add optional text to direct the transition. Check the selected duration and any charge, then generate. Review whether the door actually opens and whether the ending matches your intended composition.
- For a usable clip that needs continuation, select Extend. Drag an image from a board or upload one as the extension’s destination keyframe. The Extend guide also allows leaving the endpoint blank for an open-ended continuation.
- Add extension text if you need a particular next action, or leave it blank for the documented continuation option. Submit, then inspect the join and new ending. The tutorial places saving and sharing under the video’s three-dot menu.
Decide whether the extension earns its place
A longer clip is only useful if the additional seconds have a job. After the door opens, perhaps the camera should pause to let the viewer understand the entrance. That is a different request from continuing motion indefinitely.
Before extending, inspect the accepted clip’s ending. If the frame has already lost a glass panel or altered the doorway, extending carries an unsuitable result into the next decision. Return to the last acceptable point or rethink the shot instead of using length as the measure of progress.
| Current result | Useful next decision | Why |
|---|---|---|
| Action works but ends too abruptly | Add a clear settling beat | The extension has an editorial purpose |
| Geometry has visibly changed | Repair or replace the source segment | More time does not establish fidelity |
| A new viewpoint is needed | Consider a deliberate new shot | An edit may communicate the change clearly |
| The current ending already fits | Keep the shorter clip | Extra duration adds no required information |
These are production judgments, not undocumented Luma restrictions. They help you stop an extension chain when it ceases to improve the finished sequence.
Keep audio and reframing as separate checks
The platform overview describes native audio in some models. Ray3.14’s FAQ specifically says it does not generate native audio; it documents Add Audio after a generation completes. Plan sound as a separate operation for that model.
For the greenhouse scene, decide whether the edit needs ambient sound, music or silence behind narration. Evaluate that need after establishing a usable visual.
Decide the destination aspect ratio early. Although the platform overview describes reframing, Ray3.14’s guide and FAQ say that model cannot use Reframe. Do not base a Ray3.14 plan on that finishing option. If you crop in an editor, check that the door and the path remain visible.
No universal credit price is supplied here. Luma’s Ray2 documentation notes that added features affect usage. Review the selected operation’s current charge before treating extension, audio and finishing as included in the cost of the first clip.
Preserve the branch you want to keep
Maintain a simple folder or project record containing the original start image, intended end image, prompt, model and accepted clip. Save each extension as a new candidate until it passes review. The newest result should not automatically replace the best result.
Record the creative change in each filename: door-open, door-open-settle or entrance-new-angle. Keep the model and duration alongside it so a later revision does not silently change the setup.
Finish by watching the sequence without pausing, then inspecting the joins. Check whether the viewer understands the action, whether the greenhouse remains recognizable and whether the ending prepares the next shot. A sequence that meets those requirements is more valuable than a longer sequence whose continuity steadily becomes harder to defend.
Do keyframes guarantee a seamless transition?
No. They specify visual guidance and endpoints. Review the generated motion and the final edit. If a continuous transition is not working, a purposeful cut or a simpler action can be the better creative decision, even when more generation options are available.
Choose a character workflow when interaction is the goal
A greenhouse shot is a visual asset. A companion character also needs a personality, a scenario and a reason for the next interaction. Cherrypop’s creator provides those fields alongside realistic or anime appearance choices, then offers routes into chat and media creation.
The Cherrypop video entry includes image-to-video and continuation modes in the current interface. Use it when the clip belongs to a character experience. We are not claiming that it exposes Luma’s specific keyframe controls, model durations or separate finishing options.
Cherrypop is free to start with limits; relevant features may require Premium or Cherries. Check the displayed operation cost. For a character-led project, create the character and scenario before deciding which motion should happen next.