Image to Video
Image to video is the job of turning a single still picture into moving footage: the model reads your image, takes a written description of what should happen, and predicts the frames that follow. Upload a photo on the left, say what should move, and you get an MP4 of one to thirty seconds back in one to five minutes, depending on the model.
Seven models run this route, all priced per second: Grok Video 1.5 from 25 credits for five seconds at 480p, Wan 3.0 from 35, Wan 2.2 from 60, Seedance 2.0 from 75, MiniMax H3 from 80, Seedance 2.5 from 95 and Wan 2.7 from 105. Output is an MP4 with nothing burned into it; every model except Wan 2.2 adds its own audio track.
Describe motion, not the picture
The single biggest difference between a good result and a mushy one is what you put in the prompt. The model can already see your image — describing what is in it spends the prompt on information it already has, and leaves it guessing about the part it cannot see, which is what should happen next.
Compare two prompts for the same photo of a woman on a park bench. "A woman sitting on a bench in a park, autumn leaves, warm light" tells the model nothing it did not already know. "She turns her head towards the camera and smiles; leaves drift past in the foreground; the camera pushes in slowly" tells it what to generate.
- Name one camera movement: push in, pull back, pan left, tilt up, orbit. Two at once usually cancel each other out.
- Name one subject action. A list of five actions in eight seconds produces a clip where none of them completes.
- Say what is in the background that should move — water, cloth, hair, smoke, leaves, traffic. Ambient motion is what makes a clip read as footage rather than a moving photograph.
- Skip style words like "cinematic", "4K", "masterpiece". They come from still-image prompting and do nothing here; the look is set by the image you uploaded.
The second image: ending on a frame you choose
The optional second upload sets the last frame, and the model interpolates between the two. Instead of guessing where the motion should end up, it has a target — the clip starts on your first picture and arrives at your second.
This is the reliable way to get a specific result rather than a plausible one. Two shots of the same scene from slightly different angles give you a controlled camera move. A closed hand and an open hand give you the gesture. Two product shots give you a turn.
It needs the two images to be relatable: the same subject, the same lighting, the same general framing. Two unrelated pictures produce a morph, not a movement.
What the image needs to be
PNG, JPG or WebP, under 10 MB, with every side between 360 and 2000 pixels. The page checks all of this in your browser before anything uploads, so a file that is too large or too small is rejected immediately rather than after a wait and a charge.
Beyond the hard limits, sharpness matters more than resolution. A crisp 800-pixel photo animates better than a soft 1900-pixel one, because the model extends the detail it can find — and if the detail is already smeared, it extends the smear. Heavy JPEG artifacts, motion blur and aggressive phone night-mode processing all show up amplified in the output.
| Setting | Limit | Why |
|---|---|---|
| Formats | PNG, JPG, WebP | Checked by file type before upload |
| File size | Under 10 MB | Vendor limit |
| Each side | 360 to 2000 px | Below 360 there is too little detail; above 2000 is rejected |
| Length | 1 to 30 seconds by model | Drift grows with length; 5 holds a face better |
| Output | 480p to 1080p MP4 | No watermark; audio on every model but Wan 2.2 |
What one clip costs
Image to video is billed per second of output, not per job, so a short test costs a fraction of a full-length take. Credits are charged when the job starts and refunded automatically if it fails.
| Model | Smallest size | 5 seconds |
|---|---|---|
| Grok Video 1.5 | 5 cr/s at 480p | 25 credits |
| Wan 3.0 | 7 cr/s at 480P | 35 credits |
| Wan 2.2 | 12 cr/s at 480p | 60 credits |
| Seedance 2.0 | 15 cr/s at 480p | 75 credits |
| MiniMax H3 | 16 cr/s at 768P | 80 credits |
| Seedance 2.5 | 19 cr/s at 480p | 95 credits |
| Wan 2.7 | 21 cr/s at 720p | 105 credits |
Start at 480p, then commit
Generate at 5 seconds and the smallest size first, every time. On Grok Video 1.5 that is 25 credits and about a minute; on Wan 2.2 it is 60 credits and about two minutes. It answers the only question that matters at that point: did the model understand the motion you asked for? If it did not, the prompt is wrong and a 720p version of the same prompt will be wrong more expensively.
In our test a 1280×720 still came back as an 832×480 clip at five seconds. The framing and the motion at 480p are the same ones you will get at 720p — only the detail changes — so the cheap pass is a real preview, not a different result.
When not to use this
- Generating a scene from scratch. Every clip here starts from a picture you upload. If you have no image, use text to video instead.
- Long takes. Faces and fine detail drift the further the model gets from your first frame, and no prompt prevents it — even where a model allows 30 seconds. Generate several short clips and cut them together.
- Keeping a face exactly recognisable. Over five seconds it usually holds; over eight it often softens into someone adjacent. Treat a person’s likeness as approximate.
- Readable text. Signs, labels and captions in the source image turn to nonsense as soon as they move. Add text afterwards in an editor.
- A specific soundtrack. The models that add audio invent it from the scene, and Wan 2.2 returns a silent MP4. Add music or voiceover yourself.
- A photo of somebody who has not agreed to it. Animating a real person’s picture without their say-so is not what this is for.
Questions people ask
How do I turn an image into a video?
Upload the picture in the editor on this page, write one sentence describing what should move, pick a model, a length and a size, and press Create. It takes one to five minutes depending on the model and returns an MP4. The cost is shown on the button before you press it — 25 credits for a 5-second 480p clip on Grok Video 1.5, 60 on Wan 2.2.
What does image to video actually do to my photo?
It does not move your photo around. The model generates every frame after the first one, using your picture as the starting point and your prompt as the instruction. That is why the output can show things the photo never contained, like the far side of a face as it turns, and why fine detail changes as the clip runs.
How long can the video be?
From one second on Grok Video 1.5 to 30 on Wan 3.0 and Seedance 2.5; Wan 2.2 offers 5 or 8. Five is the better default: the further the model gets from your starting frame, the more the subject drifts, so an 8-second clip of a person is noticeably less stable at the end than a 5-second one of the same shot.
How much does image to video cost?
Per second of output, and the rate depends on the model and size: 5 credits a second on Grok Video 1.5 at 480p, 12 on Wan 2.2, 19 on Seedance 2.5, up to 79 for Seedance 2.0 at 1080p. A 5-second 480p clip is 25 credits on Grok and 60 on Wan 2.2. Credits come off when the job starts and are refunded automatically if it fails.
Is there a watermark on the video?
No. The output is a plain MP4 with nothing burned into it, at 480p or 720p depending on what you picked. It has no audio track.
Why did my clip come out blurry or warped?
Usually one of three things: the source image was soft or heavily compressed, so the model had smeared detail to extend; the prompt asked for several actions at once, so none of them resolved; or the clip was 8 seconds when the subject needed 5. Try a sharper image and one clear instruction at 5 seconds before spending more.
What is the second image for?
It sets the final frame. With two pictures the model interpolates between them, so the clip begins on the first and ends on the second — the way to get a specific movement instead of a plausible one. Both images need the same subject and lighting, or you get a morph.
Can I use it on a photo of a real person?
Only with their agreement. This is the one limit on this page that is not about output quality, and it applies regardless of where the photo came from.
Related tools
Checked 20 September 2026.