Take a photo you already have and bring one thing in it to life: steam rising from a cup, a candle flame, ripples on water. Describe that single movement to an image-to-video model, tell it everything else must hold still, and loop the result into a quiet moving picture.
A cinemagraph is a photo with one thing alive in it: steam rising from a cup, a curtain breathing in a window, a flame in an otherwise frozen room. The charm is the contrast, and it used to take a tripod, a masking job in video software and an afternoon. Image-to-video models change the starting point. You hand them the photo, describe the single movement you want, and they animate from there.
The classic method starts at the shoot. You film a locked-off video, pick the frame that works, mask the one region that should move and loop it. It produces beautiful results, but it only works if you planned the shot in advance. A photo you already have, a product shot from last spring or an image you generated yesterday, cannot become a cinemagraph that way, because there is no footage behind it.
Image-to-video flips the order. The still comes first, and the motion is added afterward by describing it in a sentence. That also means the photo can come from anywhere: your camera, your phone, or a text-to-image model running in the toolkit. Picasso IA runs 502 models, and the video ones sit in the text-to-video category, 121 models strong, many of which accept a starting image. The collection lets you browse them by category, and new accounts get free credits to try a first clip before paying for anything.
Cinemagraphs earn their place where a full video would be too loud and a still would be too quiet: a website header, a newsletter banner, a product page, a phone background. If the look appeals to you for screens, the phone and desktop wallpapers page covers that use, and the social media posts page covers feeds.
Not every part of a photo makes a good cinemagraph. The elements that work best have motion that repeats on its own and has no obvious start or finish, which is exactly what a loop needs. One-time actions, like a person walking through the frame or a door swinging open, fight the format, because the clip has to end where it began.
| Element | Prompt wording that tends to work | How well it loops |
|---|---|---|
| Steam or smoke | thin steam rises from the cup and curls upward, slow and steady | Good, the motion repeats on its own |
| Candle or fire | the flame flickers gently, nothing else in the frame moves | Good, a flicker has no start or end |
| Water surface | slow ripples spread across the pond, the shore stays perfectly still | Good if the ripples stay inside the frame |
| Hair, cloth, leaves | the curtain lifts and settles in a light breeze, the room stays still | Medium, the sway needs a visible return to rest |
| Glowing light or screen | the lantern glow pulses softly, everything else is frozen | Good when the pulse is slow |
| A small gesture or blink | a slow blink, the face otherwise unchanged | Hard, faces drift over a clip |
Pick a photo that helps the model. The moving element should occupy a clear region, the background should be simple enough that a little drift goes unnoticed, and the image itself should be sharp, because the first frame of the clip will look exactly like it. If your source is soft, fix it first with the steps on the blurry photos page, then animate.
Here is the thing that surprises people. An image-to-video model animates the whole frame by default. Left alone it will add a slow camera push, a little sway in the background, a shimmer in the highlights. A cinemagraph prompt therefore has two jobs: say what moves, and say what does not.
A workable prompt reads like this: steam rises from the teal coffee pot and curls upward in a slow, steady loop, locked camera, static shot, everything else in the frame is perfectly still.
Kling v3 Video and Wan 2.7 I2V both have a negative prompt field, so use it for the things you do not want: camera movement, zoom, shaking, new objects appearing. Ray 3.2 has no negative prompt, which means the wording of the main prompt does all the steering. One more setting matters on Wan 2.7 I2V: prompt expansion is on by default and rewrites short prompts into longer ones, which usually helps, but can invent motion you never asked for. For a cinemagraph, try it off and keep the prompt tight.
Name the things that must stay still, not only the thing that moves. A line like locked camera, background does not move is worth as much as the description of the steam.
This is the shortest path from a photo to a loop you can check and publish.
Keep the first clips short. Ray 3.2 offers 5 and 10 second durations and its loop mode only works at 5 seconds, and for a quiet detail like steam a few seconds is plenty. Cost per generation depends on the model and the settings you pick, so check the pricing page rather than guessing.
The illusion of a cinemagraph lives or dies at the moment the clip restarts. A jump there, a flame that snaps back or a cloth that teleports, tells the viewer they are watching a short video on repeat, and the spell is gone. There are three ways to approach it on Picasso IA.
Ray 3.2 has a loop option that asks for a clip that cycles without a visible cut. It is limited to the 5 second duration and cannot be combined with 10 seconds, HDR or an end image, so it suits plain, small movements. Wan 2.7 I2V lets you set the first frame and the last frame to the same image, which forces the clip to return to where it started. This works well for gentle motion and badly for dramatic motion, because the model has to hurry back to the start. Kling v3 Video also accepts a start image and an end image, so the same trick is available there, with the end image requiring a start image to go with it.
Whichever you use, judge the loop by watching it repeat for a minute, not by playing it once. The eye catches a jump on the third pass. If a small jump remains, a video editor outside Picasso IA can trim a few frames at the edges.
A cinemagraph made this way is an imitation of the classic technique, and the difference matters in five places.
The best results come from small, repeating movements in simple scenes. When what you want is a real clip with story and camera motion, the turn photos into videos page is the better starting point.
It is a looping clip made from a still photo in which only one element moves, such as steam, a flame or water, while the rest of the picture appears frozen. With AI, you upload the photo to an image-to-video model and describe that one movement in a prompt. The model generates a short clip from your photo, which you then loop. It sits between a photograph and a full video.
Most photos can be animated, but some work far better than others. The best candidates have one clear element with repeating motion, like steam or a flickering light, a simple background and a sharp image. Busy scenes, soft photos and pictures where the interesting part is a face or a crowd tend to give uneven results. If a photo does not work, try a crop that isolates the element.
There is no single answer, so it helps to try the same photo and prompt on a few. Kling v3 Video and Wan 2.7 I2V both take a starting image and a negative prompt, which helps keep the background still. Ray 3.2 has a dedicated loop option for 5 second clips. All three are in the toolkit, so you can compare them side by side on one photo and keep the one that behaved best.
You cannot guarantee it, but you can improve the odds a lot. Say it in the prompt: locked camera, static shot, everything except the steam stays perfectly still. Add camera movement, zoom and shaking to the negative prompt on models that offer one. Keep the clip short, since drift grows with length. If something still moves, composite the clip over the original photo in a video editor so only your element shows.
Ray 3.2 has a loop setting that works with 5 second clips. On Wan 2.7 I2V, pin the same image as both the first and last frame, which makes the clip return to its start. Kling v3 Video also takes a start and an end image. After generating, watch the clip on repeat for a while, since a small jump at the join is the most common flaw and may need another attempt or a trim.
Short. A few seconds is enough for a quiet movement like rising steam, and longer clips give the background more time to drift and give you a bigger file to host. Ray 3.2 offers 5 and 10 second clips, with looping limited to 5 seconds. Kling v3 Video defaults to 5 seconds and can go longer, but for a loop there is little reason to. Make the movement itself long and slow instead.
Only if it is a photo of yourself, someone who has agreed, or a fictional character, and never a real identifiable person who has not. There is also a practical limit: faces tend to drift and change over a clip, so a loop of a face often looks uncanny. A safer choice is to animate something near the person, like steam from their cup or hair in a breeze, and keep the face still.
Yes, and it is a good route when you do not have the right photo. Generate the still with one of the 212 text-to-image models in the toolkit, composing the scene so one element is clearly separate from the rest. Then use that image as the starting frame of an image-to-video model and write the movement prompt. Since you control the composition from the start, you can design the scene for the loop.
Often yes, but check two things. First, the output is a video clip, so it works anywhere video does; turning it into another format such as an animated GIF is a separate step in another tool. Second, platforms differ in what video they accept, how they handle autoplay and loops, and how long the clip may be, so check each platform current requirements before you build around it.
Each generation costs credits, and the amount depends on the model and the settings you choose, such as resolution and duration. Video costs more than images, and cinemagraphs usually need several attempts, so budget for a few runs per photo. New accounts get free credits to start. The pricing page lists the current plans, and it is the only place to trust for numbers.
Ready to bring one detail of a photo to life? Open the Picasso IA toolkit, attach your still and describe the one movement you want.
Choose your language
