To understand how to create AI video, start with one finished shot: choose its purpose, prepare an image or a short scene description, request a single clear action, review the full result and export only the usable version. Trying to generate an entire story before the first shot works makes revisions harder to diagnose.
How to create AI video: define a shot you can finish
Write a one-sentence deliverable before choosing a model. For example: “A short vertical clip of a ceramic cup on a windowsill, with steam rising, for the opening of a quiet morning montage.”
This brief defines the subject, motion, framing and use. It also tells you what you do not need yet: multiple locations, dialogue or a continuous minute of footage. A successful first clip should answer the brief without requiring an explanation from its creator.
Decide three things before generating:
- The shape of the final frame, based on where you will use the clip.
- The action that must be visible during the shot.
- The details that must remain stable, such as the cup’s handle and the window behind it.
Keep text overlays outside the first generation exercise. If you need a title or a product price, add it during editing so you can correct it without recreating the footage.
Choose text-to-video or image-to-video
Use text-to-video when the appearance of the scene is still open. Use image-to-video when you already have a composition you want to animate. The second route lets you inspect the starting image before asking for motion.
| What you have | Useful starting route | Prepare first |
|---|---|---|
| A written idea, with no fixed appearance | Text-to-video | Subject, setting, action and camera instructions |
| A finished still | Image-to-video | A clean crop and a motion prompt |
| Two required compositions | A tool with first and last frame controls | Compatible opening and ending images |
| Existing recorded footage | Editing or video-to-video, if supported | Identify exactly what needs changing |
Do not upload an existing video into an image slot or assume that a reference-image field controls the final frame. Read the label for the specific task. These are different inputs, even when they appear in the same workspace.
Prepare the image before asking it to move
For the cup exercise, select a clear image with enough room above the cup for the steam. Crop it to the intended shape and inspect the handle, rim and background. Correct a visible flaw now instead of asking the video prompt to repair it while also creating motion.
Runway’s image-to-video guidance treats the image as the starting frame and the prompt as the description of motion. It also warns that visual defects in the input can become more pronounced during animation. Runway’s prompting guide
Our practical rule is to approve the still as a deliverable in its own right. If you would reject it as a photograph, identify why before spending a video attempt on it. A larger resolution will not fix a confusing composition.
Write a prompt with observable instructions
For an image of the cup, a starting prompt could be: “Thin steam rises from the cup and drifts toward the window. The cup remains on the sill. The camera stays fixed. One continuous shot.”
For text-to-video, add the missing visual information: “A blue ceramic cup sits on a wooden windowsill in soft morning light. Thin steam rises and drifts toward the window. The camera stays fixed. One continuous shot.”
These are deliberately small requests. You can judge whether the steam rises, the cup stays put and the view remains continuous. A phrase such as “make this cinematic and amazing” supplies no similarly clear acceptance test.
If you want a camera move, replace the fixed-camera sentence with one specific move. Do not ask for a stationary frame and an orbit in the same short prompt. If an action needs several distinct moments, either allow appropriate time or split it into separate shots.
A documented route through Runway Gen-4.5
Runway’s guide places Gen-4.5 in Apps View or the Video model selector. For image-to-video, upload the still to the prompt window, add the prompt, review the crop and duration, then generate. Completed outputs have download controls. Gen-4.5 requires Standard or higher; this is not a free-tier tutorial. Runway’s creation instructions
Use the shortest available duration that gives your action enough time. Set a small attempt budget and inspect the displayed charge before confirming. Those are planning choices, not a claim that a particular setting always yields the best result.
If you choose another provider, carry over the brief and review criteria, then map them to that provider’s controls. Do not carry over Runway’s entitlement or export assumptions.
When you need a specific final frame
A reveal that must end on a particular composition needs more than a starting image. Adobe documents first and last frames in Firefly Video. It notes that adding the last frame disables Motion controls, so decide whether the endpoint or a camera preset is the more important constraint. Adobe’s frame guide
Prepare an ending image that could plausibly follow the opening one. A modest change in object position is a clearer first experiment than a sudden change of room, lens, lighting and subject. The prompt can describe the journey, but it does not remove the need to inspect the generated middle.
For the cup example, a second image is unnecessary if you only want rising steam. Adding an input merely because the tool offers it can complicate a simple shot.
Review the result in three passes
Watch the clip once at normal speed without pausing. Does the action read clearly? Then scrub through the motion, paying attention to the subject’s outline and any contact with the background. Finally, inspect the start and finish as edit points.
| Problem you observe | First revision to try |
|---|---|
| The important detail is cropped out | Fix the input crop before changing the prompt |
| Too many actions compete | Keep the essential action and remove the rest |
| The camera move obscures the subject | Simplify the camera instruction |
| A stable object visibly changes shape | Recheck the input and reduce the requested transformation |
| The shot is fine but the pacing is slow | Try trimming in an editor before regenerating |
These are diagnostic priorities, not guaranteed fixes. Write down what you changed and why. If you change the image, model and prompt together, a better result will not tell you which change mattered.
Keep an accepted clip once it meets the brief. Additional attempts can be useful, but they should answer a remaining problem, such as a visibly weak ending, rather than chase an undefined idea of perfection.
Assemble and export the finished video
For a short sequence, give each shot one role: establish the scene, show the action or reveal a detail. Place the accepted clips in an editor, trim weak edges and check that adjacent shots make sense together. Add captions, titles and any audio you are entitled to use there.
Open the exported file and watch it from beginning to end. Verify the frame shape, the actual crop, the sound and any visible branding. A good preview inside the generator is not a substitute for checking the file your audience will receive.
Save the input, prompt, model name, settings and accepted export together. That record makes a later revision more manageable than trying to recreate a successful clip from memory.
Starting from an original character
For a character-led project, Cherrypop lets you create a character and access video generation. Inspect the current mode and Cherries or plan requirements before proceeding. The entry point does not imply a particular duration, free output or perfect visual continuity.
Whether your subject is a cup or a character, complete one understandable shot first. Once its preparation, revision and export steps work for you, use that process to plan the next shot.