Image-to-video is the most controllable way to make AI video. You decide the composition, character, colours and lighting in a still image, then the model adds motion. But many people are disappointed by their first results: the subject warps, the face changes or nothing moves at all. Almost always, the problem is either the source image or the prompt. This guide fixes both.
Prepare the source image
The image is the foundation. A good source image:
- matches the output aspect ratio, so it is not cropped or stretched;
- is at or near the model’s native resolution, not tiny and not enormous;
- has a clear subject with good lighting and visible features;
- leaves room for motion: if a character will walk right, leave space on the right;
- avoids heavy text, watermarks or tiny details that the model will struggle to keep stable.
If you generate the image yourself, do the hard creative work at this stage: images are fast and cheap to iterate.
Prompt the motion, not the picture
The most common mistake is describing the image again in full. The model can already see it. Your prompt should focus on:
- Subject motion: what the subject does. “She slowly turns her head towards the window and smiles.”
- Environmental motion: what moves around them. “Curtains sway gently, dust floats in the light.”
- Camera motion: one simple move. “Slow push-in.”
- Style continuity: a short reminder of the look if needed. “Soft natural light, cinematic.”
Example: for a portrait of a woman by a window, a good prompt is “She slowly turns towards the window, a soft smile, curtains swaying gently in a breeze, dust particles floating in the sunlight, slow push-in, cinematic.” It describes change, not content.
Match motion to the image
Ask for motion that makes sense from the starting frame. If the character’s back is to the camera, asking them to reveal their face forces the model to invent a face, which may not match your design. If a car is parked, asking it to race away needs a lot of new background. Small, natural motions work best: breathing, blinking, hair in the wind, a slight head movement, water flowing, clouds drifting, lights flickering.
Motion strength settings
Many image-to-video tools include a motion strength or similar setting.
- Low: subtle, living-photo motion. Very stable.
- Medium: natural movement with good stability. A sensible default.
- High: big motion, more risk of warping and identity drift.
Start low, increase gradually and stop when artifacts appear.
First and last frames
Some models accept both a first and a last frame. The model then creates motion that connects them. This is excellent for:
- controlled transitions between two compositions;
- a character moving from one specific pose to another;
- product reveals and morphing effects;
- making clips that loop, by using the same image for both ends.
Keep the two frames reasonably similar in lighting and style, and make sure the change between them is achievable in the clip length.
Avoiding identity drift
Faces may slowly change during the clip. To limit this:
- keep clips short;
- avoid big head turns and fast movements;
- use a clear, front-facing, well-lit face in the source;
- lower motion strength;
- trim the end of the clip if drift appears late.
Extending a shot
To make a longer shot, take the last frame of a good clip and use it as the first frame of the next. Keep the prompt and settings the same. Small changes can accumulate, so check each extension and consider cutting to a different angle after two or three extensions rather than continuing indefinitely.
Using your own photos
Animating your own photographs, such as landscapes, products or artwork, is a great use of image-to-video. With photos of people, consent is essential. Only animate photos of real people who have agreed, and never in ways that could embarrass, sexualise or misrepresent them. Old family photos animated for private memories are a common, harmless use; animating a stranger’s photo is not.
Examples
| Source image | Motion prompt |
|---|---|
| Mountain lake at dawn | “Mist drifts slowly across the water, gentle ripples, soft light rising, static camera” |
| Product bottle on a table | “Slow 30-degree orbit, light reflections slide across the glass, subtle background bokeh” |
| Fantasy knight portrait | “Cape moves in the wind, embers float past, the knight raises his eyes to the camera, slow push-in” |
| Rainy street illustration | “Rain falls steadily, neon signs flicker, puddles ripple, a slow pan right” |
A quick troubleshooting list
- Nothing moves: add clear motion verbs and environmental motion; raise motion strength slightly.
- Subject warps: lower motion strength, simplify the motion, shorten the clip.
- Colours change: remove style words that conflict with the image; lower guidance.
- Image is cropped: match the source aspect ratio to the output.
Building a shot from a still: a full example
Say you need a shot of a lighthouse keeper looking out to sea. First, generate a still image: a weathered man in a navy sweater at a lighthouse window, stormy sea outside, cool grey light. Try ten variations and pick the one with the clearest face and best composition, leaving space to the right of the frame where he will look. Next, write a motion prompt: “He slowly lifts his gaze towards the sea, rain streaks run down the glass, waves crash in the distance, a very slow push-in.” Generate four seeds at medium motion strength and keep the most stable one. If the last second drifts, trim it. The result is a shot that looks exactly like your design, with motion that feels natural, made with far fewer attempts than text-to-video would need.
Frequently asked questions
Should my prompt describe the image at all?
Only briefly, if it helps the model understand the subject. Focus on motion.
Can I use any image?
Use images you created, licensed or have permission to use, and get consent for recognisable people.
Why does the end of my clip look worse than the start?
Drift builds up over time. Use shorter clips or a last-frame anchor.
