AI workflows
Use image to video AI with a deliberate motion anchor
Image to video AI begins with a still and generates movement from that visual starting point. The image can establish composition and appearance, but it does not supply the unseen movement between future frames. Choose a still that can begin your intended action, then judge the whole moving result.
In this guide
- Visible scene
- Faces, clothing, light and composition are already present for inspection.
- Motion request
- Text describes the action, camera movement and progression you want to test.
- Unseen material
- Hidden hands and a face turned away still leave information to be generated.
- Existing footage
- A video source supplies a temporal performance; a single still does not.
Choose the first moment of one shot
Think of the image as a proposed opening frame. For a doorway reunion, it should show the partners before they step closer, with enough room for the gap to close. A still where they are already embracing begins a different action: settling into the hold, changing a gaze or releasing. The same characters can appear in both images without making the two starts interchangeable.
Runway's current image-to-video prompting guide separates the image's visual information from the text's direction of motion. That is a useful division of work even when you later choose another tool, provided you read that tool's actual input rules. Describe the movement that begins from the picture instead of writing a second scene that contradicts it.
Mark the opening state in your notes. For example: both partners standing, survivor nearest the open door, other partner facing inward, neither arm raised. This small record gives you something definite to compare with the generated beginning. It also exposes a common mismatch: requesting a walking entrance from an image where the person is already seated.
Read an anchor as guidance rather than a guarantee
A visible face provides a basis for comparing likeness, and an existing jacket provides a wardrobe cue. Neither guarantees that the details remain unchanged while the character turns or moves behind another person. The output still needs review. Keep separate notes for what is clearly visible in the source and what your request asks the model to infer.
This distinction is especially useful for hidden surfaces. A front-facing portrait does not document the back of the person's head. A hand concealed behind a body supplies no finger arrangement. If the proposed action reveals either, the result contains newly generated material. Decide whether that invention is acceptable for the fictional scene before treating it as a faithful depiction.
The starting image also cannot prove future timing. A close distance may support a small lean, but it does not specify how long the partners hesitate or where a camera move ends. Write those as goals for review, then assess the actual clip. Avoid treating every sentence of a prompt as an editing control with frame-level precision.
Find the space the action needs
A motion anchor should leave visible room for the action. For a hand reaching toward a shoulder, check the gap between the bodies and the crop around the relevant elbow. For a head turn, consider what will enter or leave view. An image that is beautiful as a portrait may be awkward as the beginning of a moving interaction.
Do a brief crop rehearsal without claiming it predicts the render. Imagine the final display frame and mark the subjects' likely paths. If a partner begins at its edge, a small step can make their face disappear. If both hands begin outside the image, the first contact will depend heavily on invention. A wider starting composition may give you more evidence to judge.
Keep image preparation separate from motion trials. If you need to repair the composition, approve the revised still before changing the motion request. Otherwise a better result could come from the new image, the new wording or chance, and you will have little basis for deciding which change mattered.
Distinguish a still-led shot from modifying footage
Image to video AI generates a temporal event from a still. Video modification starts with an event already recorded across frames. The second route can be appropriate when a specific performance or camera move is the material you want to transform. It is a different question from bringing a chosen portrait into motion.
Luma's Ray3 Modify documentation separates source footage, character references and keyframes. That gives a concrete example of why “video to video” should not be treated as an extra upload slot in an image workflow. The source clip provides movement to examine before modification; the still-led approach has no recorded in-between action to preserve.
| What you need to keep | Starting material to consider | What remains to review |
|---|---|---|
| A composed opening scene | One approved still | Generated action and unseen details |
| A recorded embrace performance | Permissioned source footage | How modification changes that performance |
| An intended final composition | A supported end-frame workflow | The transition between supplied states |
| The same people across new shots | Reusable identity references | Likeness and assignment in every shot |
Write one action that follows from the picture
Begin with the smallest movement that serves the scene. If the still shows two people facing each other, a shared look may be enough for the first experiment. A request to run, embrace, spin and transform introduces several places where the image stops being a strong visual anchor.
The proposed request below holds composition steady so you can compare the moving faces with the still. It is an authored example rather than a tested prompt. A fixed camera is a review choice, not a guarantee that a tool will keep every pixel stable.
A still-led gaze shot
From this starting arrangement, the partner beside the doorway slowly turns their eyes toward the other partner. The other partner makes a slight head turn in response. Keep both bodies in their existing positions and use a fixed medium view. End with both faces still visible.
Review the first frame, the change and the end
Compare the opening with the approved source. Did the shot begin in the expected arrangement, or did it introduce a different pose immediately? Then inspect the moment with the largest visible change, such as the middle of the head turn. Finally inspect the ending. A strong resemblance at the start is useful evidence only for that part of the clip.
Watch the entire sequence at ordinary playback speed as well. Pausing can help identify a face change, but it may hide sliding movement or an abrupt acceleration. Note whether the camera, background or clothing changes without a reason in the brief. Decide which changes are acceptable fiction and which break continuity.
When comparing revisions, keep the source still fixed and alter one motion instruction. Save the earlier request beside the revised one. If you change the still, label that as a new setup rather than presenting it as evidence that different wording solved the original problem.
Return to the input when the action exceeds it
If the action repeatedly exposes a region the image never showed, consider an input that documents more of that region or a shot that does not reveal it. If preserving a particular embrace is the main requirement, permissioned footage may be the more relevant starting point. If likeness across several compositions matters, construct a reference set before animating the shots.
End the trial with a specific decision: keep this motion, revise this action, replace this starting image or change the input family. That decision is more useful than declaring a tool good or bad from one clip. It also gives the next experiment a clear question.
Zombie Couple currently offers a local casting preview and reference-film playback; generation and checkout are disabled. The external workflows here come from official documentation, with no paid render test. Use the linked brand guides when you are ready to check a particular tool's controls.