Turn one photo into a short AI video

What image-to-video actually does

You give a model a single still and a short description of how it should move. The model treats your photo as the first frame and paints the next few seconds around it: a slow push in, hair moving, steam rising off a coffee, leaves in the wind. Nothing in the frame is redrawn from scratch, so the person or product in your photo stays recognizable.

This is different from generating a video from a text prompt alone. There you get whatever the model imagines. Starting from your photo means the look, the lighting, and the subject are already locked, and you are only directing motion.

Five seconds sounds short. For a loop on a product page, the opener for a reel, or a moving portrait, it is usually plenty, and it keeps each attempt cheap enough to try a few times.

What you need before you start

Figgshield is an MCP connector that your AI agent calls to make media. You add the server once at https://api.figgshield.ai/mcp, and from then on you ask your agent in plain language and it calls the tools for you.

It runs inside the tools you already use: Claude desktop and Claude Code, Cursor, Windsurf, and ChatGPT in Developer Mode on a paid plan. Any MCP-capable agent works the same way.

There is no free tier. A $5 Pro trial runs for seven days and comes with 150 credits, and if you cancel before day 7 you pay nothing more. One credit balance covers both images and video, so you are not juggling two wallets.

Point Figgshield at your photo

There are three ways to hand your still to the video model, and you pick based on where the photo lives.

• Already generated it with Nano Banana 2 in the same agent? Pass its id with start_from_generation. This is the cleanest path: image first, then video from that exact result.
• Have a file on your machine? Ask for an upload slot with create_image_upload, send the bytes to the URL it returns, then reference it with start_upload_id.
• Photo already public on the web? Point at it with start_image_url.

Write a motion prompt that behaves

The most common mistake is asking for too much. A single photo has one point of view, so big camera moves and new scenery tend to warp. Describe one or two small things: a gentle push in, a breeze through the trees, the subject blinking and turning slightly toward the light.

Name the motion, not the edit. “Slow dolly forward, leaves drifting, soft flicker of sunlight” gives the model a job it can do. “Make it cinematic” does not.

You animate the still with generate_video using Kling 3, or Kling 3 Turbo when you want the result back faster. Video is queued rather than instant, so your agent polls get_generation until the clip is ready. A reference image you send is used in memory and never stored.

Three shots worth trying with one portrait

If you only have a single headshot, these three prompts each read very differently on screen, and none of them fight the still.

• The living portrait: subject breathing, a slight head turn, eyes catching the light, hair moving a little. Good for an about page or a profile header.
• The reveal: a slow push in toward the face while the background falls gently out of focus. Turns a flat photo into an intro shot.
• The atmosphere: steam rising, dust in the light, a curtain moving in a slow breeze. Motion lives in the environment while the subject stays still, which almost always looks clean.

What a clip costs and how long it lasts

A five-second clip is 35 credits. Generating the still first with Nano Banana 2 is 8 credits at 1K or 2K, or 16 at 4K. So a photo you shot yourself plus one animation is 35 credits, and a still you generate then animate is around 43.

With the 150 credits in the trial, that is roughly four clips from your own photos, or about three full runs if you also generate the still each time. We bill the models at cost and add no markup, so the credit price is the model price.

Every finished file sits on a private link for seven days, then it is deleted. Download the clip when it is ready. Nothing is kept for you after that, so treat the link as temporary storage, not a library.