AI image to video starts with a picture you can already approve. Upload that image to an image-to-video mode, describe one visible movement, choose a compatible frame and duration, then review the downloaded clip before putting it in an edit. A slow, readable action is a better first assignment than a complete scene with several cuts.
This six-step workflow develops an original example: a blue paper boat drifting across a shallow stone basin. The prompts and troubleshooting cases below are instructional designs, not published model-test results. They show what to request, what to inspect, and how to change the next attempt without guessing which variable mattered.
1. Choose an image that can support the movement
The starting picture needs room for the action. Our boat sits left of center, points right, and has open water ahead of it. Its folded edges are distinct. The basin rim remains visible, giving the viewer a stable reference for movement. A tightly cropped boat against a featureless background would make it harder to tell whether the subject moved or the entire image slid sideways.
Inspect the image at its actual pixel size. Look at the bow, the fold where paper meets water, and the shadow beneath the boat. Repair a broken fold before animating it. A video model can turn a tiny still-image defect into an obvious changing shape. Choose an original photograph, your own illustration, or another image you have permission to transform and publish.
Separate required details from flexible details. In this example, the boat must remain blue, keep one pointed bow, and stay afloat. The exact shape of each small ripple can vary. This distinction gives the review a clear purpose. Otherwise an attractive but incorrect result can quietly replace the intended shot.
For a person, the equivalent required details might be facial identity, clothing, and a visible prop. Keep a record using the character consistency checklist before requesting large turns or gestures.
2. Select image-to-video and confirm the crop
Choose a mode that actually accepts a starting image. Uploading an image to a general project library is not always the same as assigning it as the generation input. Confirm that its thumbnail appears in the relevant input slot. If the interface distinguishes a first frame from a reference image, choose deliberately: those inputs can serve different purposes.
Check the preview after selecting the output aspect ratio. A landscape photograph placed into a portrait frame may lose the water ahead of the boat. Reframing the source first is often more reliable than asking the model to invent missing space. Keep the bow and the intended stopping point comfortably inside the composition.
Record the provider, selected model, mode, duration, aspect ratio, and any resolution setting. These are part of the experiment, not administrative trivia. A later attempt in a different mode does not isolate the effect of a revised prompt. An upscaled export also does not prove the underlying generation used that resolution.
Runway's image-to-video guidance distinguishes the picture's visual information from the prompt's motion instructions. That is the useful starting principle here. If you have no approved image and need to invent the scene itself, use the separate text-to-video workflow.
3. Write one action and one camera instruction
Begin with a restrained prompt:
Locked camera, medium close view. The blue paper boat drifts slowly from left to right across the water. Small ripples spread behind its stern. The boat remains upright and keeps its folded paper shape throughout the continuous shot.
This contains a subject action, an environmental consequence, and a camera decision. It does not redescribe every color in the source picture. The still already supplies the basin, boat, and lighting. Repeating a long description can distract from the change you actually need.
A locked camera is particularly helpful for a first attempt because the basin rim provides a stable background. You can judge the boat's movement against it. Once that shot is usable, make a separate camera variant:
The camera slowly tracks right beside the drifting blue paper boat, maintaining the same distance and viewing angle. Gentle ripples trail behind it. The boat stays upright in a continuous, unhurried shot.
Do not combine a tracking move, orbit, zoom, rack focus, and sudden acceleration merely because each sounds cinematic. Every additional instruction introduces another relationship the clip must preserve. More elaborate camera work belongs in its own planned shot. The video prompt examples show how to separate these choices.
4. Generate a short attempt and review the whole timeline
Choose a supported short duration that can contain the action at a natural pace. The example does not require the boat to cross the entire basin. A small displacement can communicate drifting without forcing a rapid or physically strange journey. Use the provider's current controls rather than putting unsupported parameters into the prompt and assuming they will be honored.
Watch once at normal speed with the sound off. Does the shot read immediately? Then inspect its beginning, middle, and end. Pause around changes in direction or shape, not just at evenly spaced thumbnails. A perfect first frame cannot compensate for a collapsing bow halfway through the clip.
Use this review sheet before spending another generation:
| Check | Pass condition for the boat shot | If it fails |
|---|---|---|
| Identity | One blue paper boat with stable folds | Reduce action; improve the source image |
| Movement | Boat moves relative to the basin | Make the camera stationary and name the direction |
| Contact | Boat remains on the water surface | Simplify motion and inspect contact shadows |
| Background | Basin rim stays coherent | Remove camera movement or use a wider source |
| Edit point | A usable interval has stable entry and exit | Trim the interval or regenerate the ending |
These are proposed acceptance conditions. They are not claims that a particular model achieved them. Save failed attempts with their prompts so a teammate can understand the revision. For a larger sequence, keep the intervals and dependencies in a storyboard.
5. Revise the cause instead of adding adjectives
Suppose the boat stays still while the water texture slides. Adding “realistic, beautiful, high quality” does not explain the missing relationship. Revise the action to identify the moving object and stable reference:
The camera is stationary. The stone basin remains fixed in the frame. The blue paper boat gradually travels a short distance toward the right side of the basin, leaving a small wake behind its stern.
If the boat turns into a different vessel, reduce the size of the requested change. Ask for a gentle bob rather than a full rotation. A large orbit can expose surfaces the input never showed; the model must invent them. Consider supplying another approved input or choosing a shot that does not depend on unseen geometry.
When ripples look excessive, simplify the environment. Request calm water and a faint wake. If that still fails, a restrained image animation or a real filmed insert may be a better production choice. Generative motion is useful when it solves the shot, not when it consumes every revision hour.
Change one major variable per attempt. Keep the source, duration, model, and crop fixed when testing a motion rewrite. Once you choose a usable motion, evaluate the camera variant separately. This makes the result interpretable even when outputs remain stochastic. For products, apply the stricter product identity checks before accepting an attractive camera move.
6. Export an interval that survives the edit
Download the actual video file and review it outside the generation page. Confirm that its aspect ratio, playback speed, duration, and audio behavior match your editing timeline. A preview can hide a watermark, a short frozen ending, or an unexpected audio track. Keep the original download alongside the trimmed version.
Leave a little usable footage before and after the intended cut when possible. Those handles let an editor adjust pacing without regenerating the entire shot. Avoid trimming exactly at the frame where a shape begins to wobble; inspect several frames around the proposed cut. A tiny defect can become much more noticeable when the neighboring shot creates a comparison.
For a silent insert, mute unwanted generated audio explicitly. Add licensed or original sound in the edit if needed. Do not assume a video model's native audio includes the exact ambience, language, or timing required by your scene. Mouth movement and approved speech need their own lip-sync review.
A useful delivery folder contains the approved source image, original download, selected interval, prompt, settings, and a short note such as “use middle section; ending rejected because bow deforms.” Name the take by revision rather than “final-final.” This is what makes a small experiment reusable during production.
Decide whether to animate, reframe, or film
Not every still needs generated motion. A gentle pan over a photograph can work for an explanatory video when the subject itself should remain unchanged. It is easier to preserve exact text and logos using ordinary editing than by asking a generator to redraw them continuously. If the goal is simply to reveal a diagram, a conventional crop or zoom may already satisfy it.
Use AI image to video when movement adds meaning: steam rising from a cup, a fictional scene coming alive, or a controlled atmospheric insert. Prefer a real shoot when viewers need evidence that a product works, a place looked a certain way, or an event actually happened. A generated image becoming a generated clip does not turn it into documentary evidence.
Before approving the boat example, write one sentence explaining its role: “This three-second insert gives the narrator a quiet visual pause before the next idea.” That purpose helps decide whether the movement is sufficient. The shot does not need to win a model showcase; it needs to work at its intended size, duration, and point in the edit.
For a multi-shot story, plan those roles in the script-to-storyboard workflow. Use a canvas to arrange inputs, notes, and selected takes; generation and final video editing still happen in tools designed for those jobs.
