Image-to-video is the most controllable way to make AI video. A practical guide to choosing the right photo, describing motion and camera moves, and fixing the usual warping and morphing.
Updated
Open all four studios free — no card, nothing to install.
Ask an AI model to make a video from a sentence and you roll the dice on everything at once: the look, the light, the framing and the motion. Sometimes it's lovely. It's almost never what you pictured.
Start from an image instead, and three of those four dice are already settled. The picture fixes the look. The model only has to add movement. That is why image-to-video has become the working method for most people making AI clips for ads, Reels and product pages — and why the AI video tools of 2026 put so much weight on it, alongside vertical output and better camera control.
Here's how to get clean results.
You give the model a still image and a short description of motion. It treats your image as the first frame and predicts the frames that follow. Anything visible in the image, it can move convincingly. Anything not visible — the back of a head, the inside of a closed hand, what's beyond the edge of the frame — it has to guess. Most strange results come from asking for a movement that needs a lot of guessing.
Hold on to that idea and the rest of this guide follows from it.
A good starting frame has:
The image can be a real photo — your product, your café, a portrait with consent — or one you generate. In Vision Studio, a common flow is to create the still on the Image desk, pick the best one, and send it to the Video desk as the reference. Our image prompt formula covers getting that still right.
The biggest mistake is describing the picture all over again. The model can see the picture. Tell it what changes.
A simple three-part formula:
Subject motion + camera motion + pace and mood
| Part | Examples |
|---|---|
| Subject motion | steam rises · hair moves in a light breeze · she turns her head and smiles · waves roll in · the fabric ripples |
| Camera motion | slow push-in · gentle pan left · orbit slightly to the right · static camera · slow tilt up |
| Pace and mood | slow and calm · smooth, cinematic · natural, real-time speed |
Put together:
You only need a handful of terms, and using the standard word helps.
One camera move per clip. "Pan left while zooming in and orbiting" produces soup.
Remember the guessing problem. The less the model has to invent, the better it looks.
| What you see | Likely cause | Try this |
|---|---|---|
| Face changes or melts | Too much head or camera movement | Smaller turn, static camera, closer and sharper starting image |
| Extra or fused fingers | Hands partly hidden in the image | Choose an image with hands fully visible or out of frame |
| Text or logo wobbles | Models can't hold fine lettering | Remove text from the image; add it in the editor |
| Background "breathes" or shifts | Vague prompt | Say "static camera" and "background stays still" |
| Motion is too fast or rubbery | No pace given | Add "slow, smooth, real-time speed" |
| Nothing really happens | Prompt describes the picture, not motion | Rewrite with a clear verb for the subject and the camera |
Change one thing at a time between attempts, exactly as with image prompts. If you change the image, the prompt and the model together, you learn nothing from the result.
A generated clip is a few seconds long. That's fine — so is a shot in any advert. Plan a short sequence:
Four clips of four seconds, cut together with music, make a tidy 15-second ad. To keep them matching, generate all the starting images in one session with the same style words, so colour and light stay consistent.
The raw clips are ingredients. In Motion Studio, put them on the timeline, trim the weak first and last half-second of each (the start and end of generated clips are often the least stable), add your title and captions as real text, then lay in music or a voiceover. You can generate a narration on Vision Studio's Audio desk; our AI voiceover guide explains how to script it so it sounds natural.
Take one good photo of a cup of chai by a window. Generate three clips from it:
Watch them side by side. In ten minutes you'll understand more about motion prompts than from a week of reading, including this article.
Text-to-video creates the whole scene from your description, so the look changes every time. Image-to-video starts from your picture as the first frame and adds motion, so the subject, colours and composition stay as you chose them.
Only the movement. The image already shows what things look like. Describe what the subject does, how the camera moves and how fast, for example: 'Steam rises slowly from the cup, camera pushes in gently.'
Large or fast movements force the model to invent details it can't see in the image, such as the other side of a face or hidden fingers. Smaller, slower motion and a clear starting image reduce it.
Most models produce a few seconds per generation, typically around five to ten. Longer videos are built by generating several clips and editing them together.
Only with their permission, and label the result where platforms require it. Never animate a real person to make them appear to say or do something they didn't.
| Subject changes into something else |
| Asked for too large a change |
| Break it into two clips |
