What image-to-video is
There are two ways to get video from AI. The first, text-to-video: you describe the scene in words and the model invents it from scratch. The second, image-to-video: you give it a finished image and the model builds motion from it.
The second almost always gives a more predictable result, and here is why. When you describe a frame in words, the model decides for itself how people, lighting and setting look. When you give it a photo, all of that is already set. It only has to work out what happens a second later.
Which leads to the main rule.
The rule: describe the motion, not the picture
The most common mistake looks like this. Someone uploads a photo of a girl in a café and writes the prompt: «a girl with dark hair sits at a café table, a cup of coffee on the table, warm light».
The model already sees that. You have just retold the content of the picture and said nothing useful.
A good prompt describes what is not in the photo: the motion.
The girl slowly lifts the cup and takes a sip, steam rises, a car passes outside the window, the camera slowly pushes in.
That is all. Appearance, lighting and setting come from the photo. Your job is to say what comes alive.
What is more, describing the appearance when there is a photo hurts: the text starts to conflict with the picture, and the model looks for a compromise between what it sees and what it reads. As a result the face drifts.
How to set up motion
Motion in the frame comes in two kinds, and you need to describe both.
What moves in the scene. Hair in the wind, steam over a cup, leaves, water, smoke, blinking, a slight turn of the head, a hand movement.
What the camera does. This is where the language of cinematographers works, and models understand it:
- slow push-in (dolly in),
- pull-out (dolly out),
- sideways track (truck),
- panning turn (pan),
- orbit around the subject (orbit),
- camera rise (crane up).
The combination «gentle motion in the scene plus slow camera movement» almost always looks good. The combination «sharp object motion plus a fast camera» almost always falls apart.
The rule of less motion
The more you ask for, the higher the chance the model breaks the image.
Ask for the minimum. One turn of the head instead of turning the whole body. A blink instead of a change of expression. A slow push-in instead of an orbit. Modest motion on a beautiful still frame looks more expensive than sweeping motion with broken anatomy.
If the same hero should appear in several clips, you will need character consistency. This is especially true for people. The face is where the viewer’s brain notices the smallest falseness. Five seconds of a person simply breathing and blinking work. Five seconds of them getting up and walking usually do not.
A special case: old family photos
One of the most common and most touching scenarios. It works, but it needs preparation.
Fix the picture first. The model reads exactly what it sees. If the source has scratches, grain and a blurry face, you will get moving scratches and a drifting face in the video. Before animating, run the photo through an editing model: remove damage, sharpen, upscale if needed.
Ask for minimal motion. A slight smile, a blink, a barely noticeable turn of the head, a very slow camera push-in. Do not ask the person to stand up, turn around or speak.
Be ready to discard. Faces in old photos are usually small and unclear, so it will take more attempts than usual. Run drafts at low resolution.
And an honest warning. The result almost always comes out similar but not identical. It is not your grandfather, it is an AI interpretation of his photo. For many people that is touching, for some it is unsettling. Worth knowing in advance.
Why the model may simply refuse
A separate situation that catches people off guard: you upload a photo of a person, press generate and get a refusal.
It is not a glitch. It is a moderation filter. Model developers fear their tool will be used to make deepfakes, so they build in protection that triggers on realistic photos of people. The model cannot check whether the photo is yours or someone else’s, so it refuses just in case.
Developers’ policies differ, and that is the key point. For example, Seedance 2.0’s description states directly that it is not meant for photos of real people: the model animates objects, animals and generated characters very well, but it will hit the filter on a real face. The same photo may go through fine in Veo, Kling or Hailuo.
What to do:
- Switch the model. The simplest and most effective. A refusal from one does not mean a refusal from all.
- Do not try to bypass the filter. Rewordings like «this is not a real person» do not work and risk getting the account blocked.
- Remove triggers from the prompt. Sometimes the refusal is caused not by the photo but by the wording: names, celebrities, speech.
That is why having several models in one window saves time: you do not need five accounts to find out which one will agree.
Which model to choose
KADR prices, 6-second clip, 1080p, sound off, October 2026.
| Model | Price, rub. | Strength |
|---|---|---|
| Google Veo 3.1 Lite | 51 | the cheapest, for drafts and simple scenes |
| Seedance 1.5 Pro | 73 | gentle motion, portraits, bulk runs |
| Kling Video O1 (720p, 5 s) | 98 | pronounced camera movement |
| Google Veo 3.1 Fast | 101 | good quality for little money |
| Kling v3.0 | 112 | people, complex dynamics |
| Seedance 2.0 | 572 | flagship, stable character |
Kling Video O1 works only at 720p and only does 5 or 10 seconds, so its price is for 5 s (10 s cost about 196 rubles). This is not a public offer.
Two special cases. If the person in the photo should repeat a specific movement or dance from your video, take Kling Motion Control: about 13.4 tokens per second at 720p, 10 seconds come to about 212 rubles. If the photo should speak or sing to your own audio, take OmniHuman 1.5: 10 seconds cost about 308 rubles.
The full cost breakdown for every model is in the article on how much a minute of AI video costs.
Where to save
Resolution. The same clip on Seedance 2.0 costs 109 rubles at 480p and 572 rubles at 1080p. More than a fivefold difference. Tune the motion at low resolution and render the final at high.
Sound. If the person in the photo does not speak, you do not need sound from the model, and it nearly doubles the price. Switch it off.
Duration. Start with short clips. If the motion does not work at 5 seconds, it will not get better at 10, only more expensive.
Five mistakes
- Retelling the photo in the prompt. The model sees the picture. Describe motion, not content.
- Too much motion. The stronger the motion, the higher the chance the anatomy breaks.
- A poor source. A blurry or small face going in means a broken face coming out.
- Describing appearance when there is a photo. The text conflicts with the image and you get a third person.
- Final quality on the first try. You pay five times more to test an idea.
A holiday idea: an AI video greeting.
If the photo is old and damaged, restore it first: restore an old photo with AI.
In short
Animating a photo is image-to-video mode: the photo as the starting frame plus a text description of motion. Do not retell the picture, describe only what should move, and ask for the minimum. Run drafts at 480p on a cheap model and switch sound off if nobody speaks.
For old family photos, fix the source first, then animate, and ask for very little motion.
