How to Make a Cinemagraph (Two Honest Methods)

A cinemagraph is a still photo with one thing moving in it, forever: steam off the coffee, hair in the wind, a neon sign flickering while the street stands frozen. Adobe defines it as a combination of a still image and a video where most of the scene is stationary while a section moves on a continuous loop.

The frozen part is the whole trick. Motion everywhere is just a video, and your eye files it away as one. One moving element inside a still frame is a small unresolved question, which is why cinemagraphs hold attention longer than either format alone.

There are two ways to make one. The real way, and the shortcut — and they don't produce the same thing.

Method 1: the real way

This is a masking technique, not a filter. Four steps.

1. Plan the isolated motion

Before you shoot, pick the one thing that moves — and check that it's genuinely separable from everything else. Steam, fabric, hair, water, a flame, a rotating sign. The failure mode is picking something that touches the rest of the scene: a person's arm moves, so their shoulder moves, so their collar moves, and now your mask has a seam that flickers.

Decide your loop type now, because it changes what you shoot:

• Bounce loop — plays forward, then reverse, then forward. The join is seamless automatically. Good for fabric shifting, hair, subtle ambient motion. Useless for anything directional: a car drives through, then drives backwards through, and the illusion dies.
• Repeat loop — plays start to end, jumps back to the start. Needs a first and last frame that match exactly. This is what directional motion requires, and it's why Adobe's advice is to start filming before the car enters and keep rolling until it fully leaves.

2. Shoot it on a tripod

Non-negotiable. Your still layer and your moving layer have to register pixel-for-pixel, and handheld drift will show up as a shimmer along every mask edge. Use a tripod, and a shutter release or a timer so pressing record doesn't nudge anything.

Give yourself a buffer of a few still seconds at the beginning and end — that's where your base still frame comes from. Around ten seconds of usable footage is plenty; you'll use two or three.

3. Mask it

Open the footage in Photoshop or After Effects. Pick the frame that becomes your still, and place a copy of it on a layer above the video. Add a layer mask to that still, and brush out the region where the movement should show through. The still layer holds the frame; the hole in it reveals the video underneath.

Then trim the underlying footage to just the motion you want looping.

The masking itself is where the craft is. Brush softly at the boundary — a hard edge announces the trick.

4. Export

GIF is the traditional export and loops forever by default, which is why cinemagraphs and GIFs got tangled together. It's also the worst option for quality: a limited color palette will band your gradients badly. Export MP4 or WebM where the platform allows it, and keep GIF for places that accept nothing else.

Method 2: AI motion from a single still

Here's the honest version, because this is where most articles start overselling.

AI image-to-video takes one still and generates motion across the whole frame. You describe what should move and the model interprets it. There is no mask — which means there's no guarantee a region stays perfectly frozen. Steer it toward isolated motion and you'll often get something very close to a cinemagraph. But "usually respects the prompt" is not the same promise as "this pixel cannot change," and if you need the second one, you need Method 1.

Be clear about what 43frames does here: it animates a still into a short clip. It is not a masking tool, and it can't freeze a region by instruction. Anyone claiming their AI tool makes true cinemagraphs from one photo is describing a mask they don't have.

So when is the shortcut the right call?

Use AI motion when the motion is ambient and diffuse — steam, water, drifting clouds, fabric, atmosphere — and "approximately still" is fine. Also when you don't have video at all, which is the common case: you have one good photograph of something that no longer exists, and there was never a tripod.

Use the real method when the frozen region has to be exactly frozen: a product with legible text, a face, a logo, anything a viewer will stare at. AI motion on a face or on typography is where the uncanny creeps in, because a slightly-wrong mouth or a wobbling letterform reads as broken instantly.

That boundary is the same one in our guide to animating a photo, which covers the general AI motion technique — a cinemagraph is the stricter, mostly-frozen case of it.

Where it's worth the effort

Cinemagraphs earn their keep in exactly the places a static image gets scrolled past and a video gets skipped: a product page hero, an email header, a social ad. The file is far lighter than real video, and it loops without asking anyone to press play.

If you're making motion for social specifically, turning photos into videos for reels covers the format and pacing conventions. For a product page, a cinemagraph is one shot in a larger set — our guide to making a product video covers the rest.

To try the AI route: upload a still, describe the one thing that should move, and keep the prompt boring and specific ("only the steam rises; everything else is perfectly still"). Vague prompts get you motion everywhere, which is the failure you're trying to avoid.

If your source is an old print rather than a fresh photo, repair it before you animate it — motion amplifies every scratch and grain artifact already in the frame. Photo restoration cleans the still first, which is the same order of operations any AI motion workflow should follow.

FAQ

What is a cinemagraph? A still where one region loops and the rest stays frozen — Adobe calls it a cross between a photo and a video.

Tripod required? For the real method, yes — the layers must register exactly.

Bounce or repeat loop? Bounce for subtle non-directional motion; repeat for anything that travels, with matching first and last frames.

Can AI do it from one photo? It gets close for ambient motion, but there's no mask — so no guarantee anything stays perfectly still.