How to turn an image into a boomerang video with AI

Key takeaways

• A boomerang from a still image is an image-to-video job. The model animates your photo; the loop comes from making the clip end on the frame it started on, or from playing it forward and then in reverse.
• Melius has a template for it. Image to Boomerang turns a static image into a short looping animation for Reels and Stories. You add the image, check the credit cost, and run it.
• The manual route pins both ends of the clip. On a video node, the same image goes into the Starting frame and the Last frame, so the model has to come back to where it began.
• One motion loops better than three. A glint, steam, a turning label, hair in a breeze. Say what moves, say what stays still, and say that it loops cleanly.
• Shoot for 9:16. Reels and Stories are vertical, so match the aspect ratio before you generate instead of cropping afterward.

Boomerang videos from a single photo

To turn an image into a boomerang video with AI, you run the photo through an image-to-video model and make the clip loop: either the clip ends on the same frame it starts on, or a short forward clip plays back in reverse. On Melius, the Image to Boomerang template does this in one run, and a video node with matching first and last frames does it by hand when you want control over the motion.

Instagram's original Boomerang stitched a burst of photos into a clip about a second long and played it forward, then backward, on repeat. That needed a camera and a moving subject. Image-to-video models remove both requirements, which is why a still product shot can now bounce. The catch is that a model left to itself rarely ends where it started, so most of the craft is in closing the loop.

What makes a boomerang loop look seamless

A classic boomerang plays forward and then reverses, so the last frame always meets the first. You never see a cut because there isn't one: the motion just changes direction.

A generated clip doesn't get that for free. An image-to-video model interpolates from the starting frame toward whatever the prompt describes, and it may end on a different pose or light than it began with. Played on repeat, that ending snaps back to the opening frame and the viewer sees the jump. There are two ways to stop it. You can constrain the ending, by giving the model the same image as its last frame, or you can generate a short forward clip and reverse a copy of it in your video editor, which rebuilds the original bounce.

Method 1: run the Image to Boomerang template

Templates on Melius are workflows that are already built and tuned. You supply the inputs and they run the steps for you. Open Explore in the app, find Image to Boomerang, and choose Try template. The run screen shows example inputs and the output they produce, and it shows the credit cost before anything generates. It's one of the presets covered in our repeatable workflows guide.

Upload your image in the input slot and run it. Image uploads can be JPEG, PNG, WebP, HEIC or HEIF, up to 64 MB and 100 megapixels. If a run fails, Melius keeps your inputs in the form so you can generate again without re-uploading.

The template is the right choice when the motion doesn't need directing: you want the photo to come alive and loop, and you'd rather not pick a model. When you do care exactly what moves, build it on the canvas instead.

Method 2: build the loop on the canvas, step by step or with Mel

Building it yourself takes a few minutes longer and puts the model and the clip length in your hands. You can also brief Mel, the Melius agent, to set this up: describe the image, the motion and the loop, and it wires the nodes for you, the same brief-to-execution workflow applied to one clip. These are the steps it takes.

1. Upload the still. Drop the image onto the canvas. It becomes a node.
2. Add a video node and match the ratio. Set the aspect ratio to the one you'll post at, 9:16 for Reels and Stories, so the model composes for the vertical frame rather than cropping later.
3. Pick a model that takes a last frame. A video node shows only the handles its selected model can use. MiniMax H3 Max and H3 Max Turbo both support first and last frames, so start there and switch models if you want a different look.
4. Wire the same image into both ends. Connect your still to the Starting frame and to the Last frame. The model interpolates from the starting frame to the last one, so the clip has to come back to your photo. A last frame only works alongside a starting frame.
5. Write the motion prompt. Name one movement and everything that holds still: "steam rises from the cup and drifts left, the cup, saucer and table stay fixed, the camera is locked, loops cleanly."
6. Keep it short. For ad motion, 6 seconds is the sweet spot: long enough to loop, quick to render. Draft at 720p and move to 1080p for the version you post.

Choosing a model for image-to-video

The model decides how the motion feels. For animating stills, Seedance 2.0 is the safer default for image-to-video, especially with people in the frame, and Veo 3.1 is strong for cinematic motion. Check the node for a Last frame handle before you use either one for the matched-frame method. If it isn't there, keep that model for the forward-and-reverse version and switch to a first-and-last-frame model to pin the ending. The current lineup is on the library of leading models page, and it changes often.

Because every model sits on the same canvas, you can wire the same still into two video nodes on two different models and run both. Comparing two models costs one extra node.

Motion that loops well, and motion that doesn't

• One accent, not a scene. A glint across a bottle, a sparkle, steam, light moving over fabric. One thing moves and everything else stays still.
• Say what stays put. Telling the model what stays still matters as much as telling it what moves. Unanchored backgrounds drift, and drift is the first thing a loop exposes.
• Always ask for the loop. Without "loops cleanly" in the prompt, the model may end on a different frame than it started on. That's fine for a one-off clip and jarring on repeat.
• Lock the camera. Pushes, pans and orbits move the whole frame, so the start and end can't match. Keep camera moves for clips that don't need to loop.
• Avoid motion with a direction. A person walking out of frame or a car driving past has to come back. Back-and-forth motion, like a nod, a swing or a pour that settles, closes on its own.

Why make boomerangs on Melius

Melius is the AI-native operating system for creatives, built on a node-based canvas. Each input, prompt and clip is a node, and edges pass one node's output into the next. For a loop, that structure is the point: the still, the motion prompt and the model are separate pieces, so you can change the prompt and re-run the clip without touching the rest.

You can brief Mel in plain language and let it build the nodes, or run the template when you just want the result. Either way every step stays visible and editable, and you make the final call on which take ships.

You own the content you create on Melius and can use, publish, and distribute it commercially, subject to the underlying model providers' terms. Melius doesn't claim ownership of your work and doesn't use it to train AI models.

Frequently asked questions

Have a still that should move? Open Melius, run Image to Boomerang on it, and compare the result with a matched-frame clip on the canvas.