Animating a photograph means asking a model to invent the frames you did not shoot. Sometimes that is a slow camera move and a bit of wind. Sometimes it is a different face. Your job is to ask for the first and refuse the second.

The two generators reviewed here are Runway and Luma AI. CapCut can also generate from an image, and it is the editor to drop the clip into afterwards. Prices and limits below come from their public pages on 14 November 2026.

Pick a photograph that can survive motion

Look for:

  • One clear subject, not a crowd of tiny faces
  • A bit of background that is allowed to move (trees, steam, cloth, water)
  • Enough pixels that you are not enlarging a thumbnail
  • No text you need to remain readable, unless you will add the text again later in an editor

Group photographs are where these models fail in public. Arms merge. A second face drifts. If the picture is a team, a family dinner or a class, do not animate it. Put it on a timeline as a still.

Crop before you upload if the subject is lost in the frame. A square or a vertical crop is easier to post than a wide scan of a whole room, and both Runway and a phone post will thank you. Runway’s Gen-4.5 help page lists 9:16 among the image-to-video aspect ratios, which is the tall phone shape.

Write the motion, not the scene

The model can already see the photograph. Runway’s Gen-4.5 help page says image-to-video prompts should describe the motion. Use three beats:

  1. Camera. “Slow push in” or “locked-off, no camera move”.
  2. What moves. “Curtain shifts slightly” or “steam rises”.
  3. What must not move. “The face stays still. The product shape does not change.”

A prompt you can reuse:

Slow push in. Soft movement in the background only. The person’s face, eyes and mouth stay still. No new objects. No change to clothing.

For a product, name the constraint harder: “The bottle label stays readable and the bottle shape does not change.” If the label melts, the take is useless even if the light is pretty. For a property, do not ask the model to add a kitchen, a view or furniture that was not in the photograph.

Avoid “make it cinematic”, “bring it to life” and “smile”. Those are how you get a different person.

Spend the cheap credits first

Both companies charge by how hard the clip is.

Runway’s Gen-4.5 page says 12 credits per second, 2 to 10 seconds, 720p, on the web, and that the model needs a Standard plan or higher. A 2-second test is 24 credits at that rate. A 10-second one is 120. Make the short one first. The free plan is 125 credits once, and it does not include every model. Confirm the model is actually available before you count on those credits.

Luma’s pricing table, for Ray3.2 image-to-video, listed a 5-second clip at 20 credits in 360p and 400 credits in 1080p. Ten seconds costs more, and the table heading says HDR multiplies the cost. Their information page says Dream Machine is not the current product. Open Ray, or whatever the app currently calls the video model, and read the credit number before you press generate. There is no free plan on the pricing page.

CapCut’s careful own article, from September 2026, says to read the credit cost shown before an image-to-video generation, and that availability varies by account and region. Do that. Do not trust a different CapCut page that calls the same feature limitless.

Generate two or three takes of the short, cheap version. Seeds differ. The third take is often the one that does not twitch.

Do it in Runway

  1. Go to runwayml.com and sign in. The Gen-4.5 help page says the model is on the web, and that you can find it from the app search or the video model selector.
  2. Choose image-to-video and upload the photograph. The help page says you can drag it onto the prompt.
  3. Paste the motion prompt. Leave the face constraint in.
  4. Set the shortest duration you can learn from, between 2 and 10 seconds. Set the aspect ratio you will actually post. The page says the image’s own ratio is the default, and that changing it can crop.
  5. Read the credit cost, then generate.
  6. Download the take that passes the checks below. The free plan watermarks video. Standard and above are described as having no watermark.

Do it in Luma

  1. Go to app.lumalabs.ai. Luma’s information page also points at an iOS app. Use whichever you have, and ignore tutorials that still say Dream Machine.
  2. Start an image-to-video generation on the current Ray model, not a randomly selected third-party model, unless you mean to. The app also hosts other companies’ models at different credit rates.
  3. Upload the still and use the same motion prompt.
  4. Pick the draft or 360p length first. The pricing page is where the 5-second and 10-second rates are written down.
  5. If the draft is good, rerender at the resolution you need. Jumping straight to 1080p HDR is how a monthly credit pot vanishes on a test.

Luma’s information page describes up to 16 keyframes and a reframe tool. Use those only after a plain generation looks right. More controls will not repair a photograph that was a bad candidate.

How to tell a bad generation

Watch at full size, then watch again paused halfway.

  • The eyes, teeth and ears should match the photograph
  • Hands, if visible, should not gain or lose fingers
  • Glasses, jewellery and labels should not melt
  • The background may move. The identity may not
  • Nothing important should enter the frame that you did not ask for

If it fails, change the prompt toward less motion before you change the model. “Locked-off camera, only the clouds move” survives more often than a request for a turn of the head.

Put the clip somewhere

A four-second shot is not a birthday film. Drop the file into CapCut, or any editor you already know, in front of the still photographs. Cut it the moment the face starts to drift, even if that leaves you two seconds. Hold the original still afterwards so viewers see the real photograph again.

Add your own caption in the editor. Do not ask the generator to write names, prices or dates. It will misspell them, and in an advert a wrong price is your problem.

For the longer structure around that shot, use the guide that matches the occasion: a tribute, a birthday, a listing or product, or a social clip.

Questions people ask

How long can an animated photo be?

On the pages checked for this guide, Runway’s Gen-4.5 makes 2 to 10 seconds and Luma’s Ray3.2 image-to-video prices are published for 5 and 10 second clips. Plan on a short shot, then edit it into a longer film if you need one.

What should the prompt say?

Describe the motion, not the photograph. Say what moves, what stays still, and how the camera behaves. “Slow push in, leaves moving, the face remains still” is more useful than “make it cinematic”.

Which tool is cheaper to try?

Runway publishes a free plan with 125 one-time credits, but its Gen-4.5 help page says that model needs a paid plan. Luma’s pricing page does not list a free plan. CapCut’s editor is free to start, and its AI image-to-video feature uses credits. Check the price in the account before you generate.