Make a Cinemagraph Test by Separating Held and Moving Regions

A cinemagraph holds two times in one frame: a region that stays still, and a region that keeps moving. The test worth running is whether those two times can share an edge. A clean loop does not prove they can. You can build a flawless cycle and still leave a shadow frozen in a place its object abandoned ten frames ago, and a viewer will feel the wrongness before they can name it.

So the work has an order. Decide what stays alive. Choose a held frame that agrees with it. Then look hard at the line between them. What you end up with is a construction you can show in a treatment—an editable file with a chosen moving region, labeled as a built image rather than a record of an unchanged moment.

That last part is not a formality. A cinemagraph is often read as a captured moment, as if the scene really was exactly this still and this alive at the same instant. It is neither. It is two moments joined, and the join is where the craft shows.

Choose a passage that can tolerate being partly held

Start from footage you are authorized to use in the treatment. Then look for two properties, in this order.

A stable viewpoint. Locked-off is ideal; a shot that reads as locked-off will do. If the camera drifts, every pixel in the frame changes, and no region can be frozen honestly. A slow handheld pan makes the frozen area slide against the live one, and no mask shape repairs that.

One legible movement. A cinemagraph asks the viewer to notice a single thing: steam off a cup, a curtain breathing, a sign turning, light moving across a wall. If several things move, you are not making a cinemagraph. You are making video with a still patch.

Then run the agreement check. Pick a candidate frame to hold, and ask whether it agrees with the rest of the image across the whole loop. Three things betray a bad choice:

• Shadows. Does anything cast a shadow that travels across an area you plan to freeze?
• Reflections. Does a window, screen, or wet surface carry the moving thing along with it?
• Contact. Does the moving thing rest on, push, or touch something outside your intended live region?

Play the candidate passage once more and watch only the areas you mean to freeze. If they change—if a cloud shifts the light, if a shadow sweeps across the desk—either redraw the boundary to include that change or choose another passage.

When the motion changes everything around it, do not force the plan. Keep more of the image live, or find a different shot. A smaller frozen area that holds is worth more than a large one that breaks.

Build the held-and-moving composite

The basic construction is straightforward. Export one frame as a still. Place the still on the lower video track and the moving clip above it, then mask the clip down to the region that should stay alive. Adobe's Premiere Pro tutorial lays out this shape: a held background built from a source frame, sitting beneath masked video, with the repetition checked afterward.

That write-up is historical, and it is not a test of your software's current controls. Treat its menu names and mask behavior as a starting point and confirm them in the version you actually use. What carries over is the structure, not the keystrokes.

Keep three things identifiable in the project: the original clip, the still you exported, and the mask shape. Name them. A test you cannot revise is not a test.

Start with a simple boundary—a rectangle, a straight-edged shape—and look at how it sits against the action before you soften anything. Feathering comes later, if at all. A soft edge can hide a visible seam. It cannot reconcile content that disagrees across the line. If a frozen shadow sits where a live shadow should be, blurring the boundary does not fix the contradiction; it only makes it fuzzy.

One more decision belongs here. The still and the live clip should agree at the point of return. Freeze the frame that matches where the moving region restarts. If the live region loops back to its first frame, hold that same frame in the still, so the two sides match at the moment the loop closes.

Follow what crosses the boundary

This is where the test is won or lost. Play the whole passage and watch the boundary itself rather than the movement. Look for shared shadows crossing the line, contact points between the live object and a frozen surface, objects entering or leaving the live area, and reflections that run past the mask edge.

The Adobe write-up names one of these directly: unwanted boundary entry is a reason to shorten the passage. That instinct generalizes. When something crosses the line, the question is not how to cover the crossing but whether the crossing should happen at all.

Take one crossing and work it through. Here is a paper walkthrough—a hypothetical scene for reasoning, not a test that has been shot or composited.

Locked-off camera. A desk in front of a bright window, and the window's light throws an object's shadow toward the camera, across the desk. The object is a small car that rolls left to right, taking its shadow with it. The clip runs three seconds at 24 frames per second—seventy-two frames. At frame 1 the car sits at the left, its shadow pointing down-left. At frame 72 the car is at the right edge, the shadow pointing down-right. The shadow stays inside your planned live rectangle for roughly the first second, then crosses its lower edge.

You draw the live region as a rectangle around the car's path and hold frame 1 as the still. Now play it. Inside the rectangle, the shadow travels with the car. Below the rectangle, the held frame shows bare desk, because at frame 1 the shadow had not reached that far. When it travels across anyway, the live shadow stops dead at the boundary and the desk beneath stays smooth and unlit. A shadow cut off along a straight, invisible line.

Try a second held frame, frame 36, about mid-desk. By then the shadow has already crossed the lower edge, so this still carries part of the shadow below the rectangle. At the loop point, the live region restarts at the left with the shadow still inside the rectangle while the held frame shows that shadow out at mid-desk, and the two sides disagree from the first frame of the loop. Keep this version in the project as your labeled failure—it is the clearest illustration of the problem. The frame-1 version is closer to working, because at the moment of return the still and the live region agree; its fault appears only mid-loop, as the shadow reaches the edge.

Neither choice makes the crossing behave, and that is the lesson. Hold an early frame and the shadow is clipped, with bare desk beneath it. Hold a later one and a shadow stays frozen where its object has already left. A held frame does not travel, so no choice of frame repairs the crossing. So repair the construction rather than the blur:

• Enlarge the live region to cover the whole shadow sweep. The shadow stays attached. The cost is that more of the image is live and less is frozen, so the nearly-still effect weakens. Check whether it still reads as a cinemagraph.
• Shorten the passage. Since the shadow stays inside the rectangle for about the first second, using only frames 1 to 24 removes the crossing entirely. The cost is a smaller movement that may no longer read as a repeating event.
• Choose a different passage. If neither the wider live region nor the trimmed excerpt holds, the shot is a poor candidate.

A repair that only softens the line, or parks the shadow somewhere plausible, is not a repair. If the shadow should travel with its car and instead stops at the edge or sits frozen behind it, the image is lying, and softer edges will not make the lie true. At that point the honest moves are: accept a larger moving region, shorten the beat, or drop the freeze and let the whole frame move—which means you have a loop, not a cinemagraph, and you should say so.

Check the return in the actual viewing copy

The loop's return is its own problem, and a companion piece on looping commercials works through it in depth. Keep it short here. Choose the briefest return that preserves the motion the treatment needs. A subtle movement slips back almost invisibly; a bold one announces its jump.

Watch several repetitions, not one. A single pass forgives a lot; by the third or fourth loop the eye catches the repeat. And separate two different faults, because they need different fixes. A spatial seam is a visible line or edge, and feathering or redrawing the boundary can help. A temporal jump is the motion snapping from end to start, and it wants a longer passage or a return the eye does not follow. They can appear together, but do not treat one as the other.

Then watch it at the size the client will see it. A boundary that holds on your editing monitor can fail on a phone, and the reverse. Export for the route you are actually delivering through, and test that specific copy rather than a preview. Animated GIF is one format the historical tutorial mentions, but it is not a universal rule: GIF's limits on color, size, and playback differ from a video file's, and platform support varies. Test the exported copy where it will actually be viewed—an attachment, a deck, a link—instead of trusting a setting you inherited.

What the test is for

The version you keep should be the shortest passage that still shows the relationship you need. If a wider live region and a trimmed excerpt keep the shadow with its car, that version beats a more frozen one that breaks. A cinemagraph's job in a treatment is not to freeze as much as possible. It is to hold two times together long enough that the viewer believes them.

And when you present it, name the failure before you show the fix. "The shadow crossed the boundary, so we widened the live area" tells a client more than a polished still that conceals the problem. A seamless still sells an effect and invites a question you may not be able to answer. A labeled construction shows you know how the effect works—and it invites the next decision instead.

Frequently asked questions

What qualities should the source footage have?

It should have a stable viewpoint—locked-off is ideal, and a shot that reads as locked-off will do. If the camera drifts, every pixel changes and no region can be frozen honestly; a slow handheld pan makes the frozen area slide against the live one, which no mask shape repairs. It should also contain one legible movement, such as steam, a curtain, a turning sign, or light moving across a wall. If several things move, the result is video with a still patch, not a cinemagraph.

How do you choose the held frame?

Pick a candidate frame and ask whether it agrees with the rest of the image across the whole loop. Check shadows that travel across an area you plan to freeze, reflections in a window or wet surface that carry the moving thing, and contact where the moving object rests on or pushes something outside the intended live region. The still and live clip should also agree at the point of return: freeze the frame that matches where the moving region restarts. If a frozen shadow sits where a live shadow should be, feathering does not fix the contradiction—it only makes it fuzzy.

What can be done when something crosses the boundary?

The paper walkthrough shows a car and its shadow crossing out of the live rectangle. Holding an early frame clips the shadow with bare desk beneath it; holding a later frame freezes the shadow where the object has already left. Because a held frame does not travel, no choice of frame repairs the crossing. The honest repairs are to enlarge the live region so the shadow stays attached, shorten the passage to remove the crossing, or choose a different passage. A repair that only softens the line or parks the shadow somewhere plausible is not a repair; if the whole frame moves, that is a loop, not a cinemagraph, and should be said.

How should the loop and final viewing copy be checked?

Watch several repetitions, not one; a single pass forgives a lot, but by the third or fourth loop the eye catches the repeat. Separate a spatial seam, where a visible line or edge can be helped by feathering or redrawing the boundary, from a temporal jump, where the motion snaps from end to start and wants a longer passage or a return the eye does not follow. Watch at the size the client will see, because a boundary that holds on an editing monitor can fail on a phone and the reverse. Export for the actual delivery route and test that specific copy; GIF is one format mentioned in the historical tutorial, not a universal rule.

What should the finished test present?

Keep the shortest passage that still shows the needed relationship; a wider live region with a trimmed excerpt that keeps the shadow with its car beats a more frozen version that breaks. Keep the original clip, exported still, and mask shape identifiable and named in the project. Present it as a built image, not a record of an unchanged moment, and name the failure before showing the fix. A labeled construction shows how the effect works and invites the next decision; a seamless still can sell an effect and invite a question you may not be able to answer.