Image to video

Image to video starts the generation from a picture you supply instead of from the prompt alone. It is the most reliable way to control exactly what is on screen, because the first frame is no longer a guess.

First frame

Upload one image and it becomes the opening frame of the clip. The prompt then describes what happens from there, not what things look like:

The woman turns her head towards the window and smiles.
Camera pushes in slowly. Everything else holds still.

Describing the picture again ("a woman in a red coat in a cafe") wastes the prompt on information the model can already see. Spend it on motion instead.

First and last frame

Upload a second image and it becomes the closing frame. The model generates the transition between the two. This is how you get a controlled result — a character moving from one pose to a specific other pose, or a scene morphing into a known end state — rather than whatever the model invents.

Both frames should plausibly belong to the same shot. Two unrelated images produce a cut or a smear, not a transition.

Image requirements

Up to 2 images (first frame, and optionally last frame).
10 MB per file.
Any common image format — JPEG, PNG, WebP.

Large uploads are resized to 2048 px on the long edge before they are sent, so there is no benefit to uploading a 6000 px master. Do use the cleanest source you have, though: compression artefacts, heavy grain, and watermarks in the input carry into every frame of the output.

Aspect ratio behaves differently here

When you supply a first frame, the output follows the shape of that image — the aspect-ratio control does not apply, because the model is expanding your frame rather than composing a new one. If you need a specific ratio, crop the input image to that ratio before uploading it.

Everything else

Model, length, resolution, audio, and reference attachments work exactly as they do in text to video, and cost is calculated the same way.

Continuing a shot

Any finished clip can seed the next one: take its last frame, upload it as the first frame of a new generation, and write the prompt for the next beat. Chaining 5-second shots this way gives you more control over a 20-second sequence than asking for 20 seconds in one pass, and it costs the same.

When results go wrong

Symptom	Usual cause
Barely moves	The prompt described the image instead of the motion. Name a specific action.
Subject warps or melts	Asked for too much movement in too few seconds. Lengthen the clip or shrink the action.
Ignores the last frame	The two frames are too far apart to interpolate. Move them closer, or split into two generations.
Output is the wrong shape	Expected — the output follows the input frame. Crop the input first.
Face drifts over the clip	Add reference images of the same person, or keep shots shorter.

Still stuck? The contact page lists what to include so we can look at the specific generation.