Image to Video AI: How to Turn a Photo Into a Video

A text to video model invents the first frame and everything after it. An image to video model is handed the first frame and invents only what follows, which is why the result looks like your photo: the person, the room and the light are yours, and the model guesses how that scene moves next. With no photo yet, start from text to video instead.

It does not restore the photo: a blurry scan comes back as a blurry clip, so sharpen first, as our animate old photos guide does. It does not know what lies outside the frame, so a camera move that reveals the room invents the room. And it holds one motion well and three badly.

The photo is the first frame, not a loose reference. Crop it to the clip's shape before you upload.

The photo decides more than the prompt does. A sharp, well lit still with one clear subject animates well on every model; a dark, busy, low resolution one fails on all of them.

Match the clip's shape to the photo: a tall photo wants 9:16, a wide one 16:9. Most models here also render 1:1, which saves a crop on a square photo. Veo 3.1 and Gemini Omni Flash 1.1 render 16:9 and 9:16 only, and Minimax Hailuo 2.3 has no shape setting, so crop the photo to the shape you want before you upload it there.

A text to video prompt describes a scene. An image to video prompt describes a change to a scene the model can already see. Writing the picture again wastes the words and sometimes fights the photo. Leave out what is already in the frame, leave out negatives, and leave out the second scene: one clip is one moment.

The order to write a motion prompt in

1. The subject's motion, first: she leans forward and blows out the candles.
2. What else moves: the flames flicker and go out, smoke curls up.
3. The camera, if it moves at all: handheld, a slow push in, or hold still.
4. The sound, on a model that renders it: a room full of people clapping.
5. The look, in two words at most: phone footage, cinematic.

The six clips on our image to video page were written this way, three on Kling V3 Turbo and three on Veo 3.1; their prompts are in the table further down. Ask for nothing sudden at the start: a jump or a fast turn in the opening frames tears the photo apart before the motion settles, and it is the commonest reason a clip ends on a different person.

These ten cover the common jobs. Each cell is read from our model catalog: the lengths, sizes and sound options the studio offers today, not a maker's advertised maximum.

• KlingKling V3 Turbo3 to 15 s · 720p, 1080p
• KlingKling 3.03 to 15 s · 720p, 1080p
• GoogleVeo 3.1 (with audio)4, 6 or 8 s · 720p, 1080p
• ByteDanceSeedance 2.54 to 30 s · 480p, 720p, 1080p
• ByteDanceSeedance 2.0 Fast4 to 15 s · 480p, 720p
• AlibabaWan 3.02 to 30 s · 480p, 720p, 1080p

Start frame is universal here; end frame and reference photos are not. When the clip has to end on a second photo, open Seedance 2.5, Wan 3.0 or Gemini Omni Flash 1.1. For a subject that should appear in a scene the photo does not show, Seedance 2.0 Fast accepts as many as nine reference photos, cited in the prompt as @Image1, @Image2 and so on.

The models differ less in whether they can animate a photo than in what they protect while they do it.

The Kling page collects prompts for people and pets, the Veo page prompts that lean on sound and light, and the Seedance page the long takes and end frames.

Length is where the models part ways, and the model table further up lists every option. Veo 3.1 stops at 8 seconds, the Kling models at 15, and Seedance 2.5 and Wan 3.0 go to 30. A photo usually wants the short end: six seconds of a smile is a moment, fifteen is a loop.

Draft at 720p, or at 480p where it is offered, and render 1080p once the motion is right. On most models the bigger size costs more per second; Veo 3.1 prices 720p and 1080p the same. For 4K from a photo, Gemini Omni Flash 1.1 renders it natively, and so do Kling 3.0 4K and Kling O3 Omni 4K, both covered in our Veo 3.1 vs Kling 3.0 comparison.

Sound follows the model. Veo 3.1, Kling 3.0, Seedance 2.5, Seedance 2.0 Fast and Wan 3.0 render sound and have it on by default, with a switch to turn it off; on Veo 3.1 and Kling 3.0 a silent clip also costs less. Gemini Omni Flash 1.1 renders sound on every clip, and the model table marks the silent ones. For a photo of a person speaking, the better route is a talking avatar, which takes the portrait and an audio track and matches the mouth to the words.

The best start frame is often one you made. A product shot that does not exist yet, a character you designed, a location you cannot visit: generate the still, judge it, then animate it. A still costs a fraction of a clip, so you pay for motion only once the frame is right.

1. Open Nano Banana Pro and write the frame as a photograph: subject, setting, light, lens, and the clip's shape.
2. Render two or three, and fix small things on the keeper with a photo edit rather than a re-roll: the image to image route takes the still and one line of change, and the Nano Banana page has frame prompts for it.
3. Open the still as the start frame in the Video studio and describe only what moves.

The first six are the real prompts behind the before-and-after pairs on our image to video page, which shows each photo beside the clip it became. The last four are new, written to the same rules and not rendered by us, so treat them as starting points. Use your own photo, keep the prompt, and change one thing at a time.

1. Open the Video studio with the start frame mode on: Kling V3 Turbo for people and pets, Veo 3.1 for products and sound, Seedance 2.5 when you need an end frame.
2. Drop the photo on the start frame, cropped to the clip's shape.
3. Write the motion line, one subject and one change, and pick the length and size.
4. Run it. Credits are held when you press and settled only when a finished clip is delivered; an animation that fails is never charged.
5. Judge the first second and the face. If either slips, calm the opening of the prompt or try the next model.

To choose between two models for one photo, run the same photo and motion line on each in the Video studio at its shortest length and smallest size, then render the keeper. Compare runs one text prompt on up to four video models side by side, each at its shortest length and smallest size in 16:9, with no charge for a failed lane; it takes no photo, so it picks a model for a motion, not for a picture.

• Hands and teeth. A hand that waves and a mouth that opens wide are where every model still slips. Keep the hands still, and use a talking avatar for speech.
• A face turning past profile. The model has never seen the back of this head; it invents one. Keep turns under a quarter turn.
• A second person entering. Anyone not in the photo is a stranger the model makes up. If they matter, put them in the still first.

Luma's Dream Machine, Pika's own app, Pollo and Veed all offer a photo to video step, as of October 2026, each inside its own editor or subscription. DreamClip runs Luma Ray v3.2 and Pika 2.2 beside Kling, Veo and Seedance in one studio, so the comparison is a model switch rather than a new account.