Image-to-Video for Beginners: Turn a Single Still Image into an AI Short Film

Text-to-video lets you describe a scene from scratch, but in much real-world work there's a more efficient starting point: an image you're already happy with. Image-to-video mode uses a static frame to lock down what the scene looks like, while your prompt only has to focus on how it moves. This division of labor makes the results far more controllable. This guide covers the essentials: when to use image-to-video, how to prepare the first frame, and how to write motion prompts.

Text-to-Video or Image-to-Video?

Which mode to choose depends on where you are in your creative process.

Text-to-video is for when you're still exploring what a scene should look like, and you're happy to let the model propose visual options.

Image-to-video is for when the frame is already decided: a product shot, a character design, a key visual for marketing, or a specific frame from a mood board. The image anchors the subject, tone, and composition, so your text can be devoted entirely to motion and camera direction.

Not all models support image input, so before preparing your assets, confirm whether the model you've chosen offers an image-to-video mode.

Image-to-video basics guide

What Is Image-to-Video (I2V)?

Image-to-video is an input mode for generative video models: the model receives a static image as the first frame, then predicts and generates subsequent frames based on a text prompt, ultimately compositing them into a short video. Unlike pure text-to-video, which relies entirely on text to "imagine" a scene from random noise, image-to-video hands the initialization conditions of the visual content to an image, and the model only needs to infer motion on top of it.

This distinction has two practical implications. The first is controllability: the subject's appearance, composition, color tone, and lighting are all locked by the first frame, leaving the model no room for free interpretation and avoiding the "every clip looks different" problem common in text-to-video. The second is a narrower role for the prompt: text no longer needs to describe what the scene looks like, only the action, camera, and details that must be preserved—making it shorter and more precise to write. The trade-off is that the model's creativity around the image itself is suppressed; if the first frame isn't high quality, the output is unlikely to improve on it.

Step One: Prepare a Strong First Frame

The starting image does most of the work, so choose carefully:

Clear subject. There should be a single obvious focal point that the motion will revolve around.
Room for motion. If the camera will push in or the subject will turn, leave space around them.
Correct aspect ratio. The frame's aspect ratio should match the format of the short film you plan to publish.
Clean details. Small flaws in a still image often become more obvious once things start moving.

If you don't have a suitable image, generate one first. Locking the frame at the static stage—for example, creating the first frame in an image generator first—separates design decisions from motion decisions, so you aren't handling both at once.

Why Does Aspect Ratio Matter?

Mainstream video generation models typically work on a fixed-resolution canvas, such as vertical 9:16, horizontal 16:9, or square 1:1. When the first frame's aspect ratio doesn't match the model's target format, the system generally handles it one of two ways: cropping content from the sides or top and bottom, or stretching the image and causing distortion. The former may cut off the subject; the latter makes the finished piece look off.

The safer approach is to get the ratio right when you prepare the first frame. If you're posting to short-video platforms, prepare a vertical image from the start; if it's for a website or a horizontal presentation, use a horizontal image. Also, most models have an upper limit on input resolution—images that are too large get downsampled, while ones that are too small may be upscaled and blurred. A good rule of thumb is at least 720p, and check the model's recommended dimensions before submitting.

Step Two: Decide on a Single Moment

A short clip can only hold one clear event. Before writing your prompt, finish this sentence: "In this clip, the viewer sees ___." Good answers are all a single beat:

The product rotates, revealing the label.
Morning light brightens as steam rises from the cup.
The character looks up from the map toward the mountains.

If your answer contains "and then," you probably need two clips.

Step Three: Write the Motion Prompt

The visuals are already defined by the image, so the prompt should describe three things:

The subject's action: what moves, and how.
The camera: static, slow push in, tracking, or panning.
Details that must remain recognizable: elements you don't want changed.

Example for a product shot:

The bottle rotates slowly to the right, revealing the front label. Static camera, centered composition. Keep the bottle shape and label colors unchanged.

Example for a character illustration:

She sets down the paper map and turns her head to look at the distant peaks. The camera pushes in slowly from behind her shoulder. Keep the green jacket and hairstyle consistent.

The image-to-video mode guide for Nano Banana 3 follows the same logic: give the viewer one focal point, choose a purposeful action, and connect the camera to the action. It also recommends mentioning any visual details you want to remain recognizable.

What Is "Camera Motion," and Does the Model Really Understand It?

In video generation, a camera instruction (camera motion) refers to simulating how a camera behaves in the scene. Common types include: static / locked-off, push in / dolly in, pull out, pan, tracking, and orbit. When you write "slow push in," the model isn't actually moving a virtual camera—it has learned from training data the pixel-change patterns that correspond to a "push in" effect (the subject gradually growing larger, the background expanding outward) and generates the frame sequence accordingly.

This means camera instructions need to be written plainly, ideally using common terms that appear in the model's training corpus. Vague phrasing like "move cinematically" usually won't trigger a stable effect. If you want the camera to stay still, explicitly writing "static camera" is more reliable than omitting it, because many models add subtle movement of their own when no camera description is given.

Step Four: Check Generation Settings

Before submitting, confirm:

The input mode is image-to-video, and the image is attached.
The format matches the platform where you'll publish the clip.
The clip length and resolution are appropriate for the use case.
The displayed cost is within your budget for this attempt.

Video generation is usually more expensive than image generation, so ten seconds spent checking settings is worth it.

Step Five: Review the Whole Clip

Watch the complete clip with your one-sentence intention in mind:

Is the action clear and readable from start to finish?
Does the camera keep attention on the subject?
Do important details change shape or color over the course of the clip?
Is the pacing suited to its use case?

Fixing Common Problems

Too much action. Cut the prompt down to one subject, one action, and one camera direction. Remove secondary motion, like passersby or changing weather.

Subject distorts or drifts. Be more explicit about the details you want preserved, and check whether the first frame clearly shows them.

Camera wobbles. Specify simple behavior, such as a static viewpoint or a slow push in, and state what it should be aimed at.

Barely moves. Write the action more concretely: "turns to the right," "lifts the cup," "the door opens."

Change only one element per attempt, so you can tell what fixed the problem.

Use Images with Clear Rights

The first frame should be something you're allowed to use: your own product photography, a generated image, or licensed material. Be especially careful with portraits of real people.

Training Data, Portrait Rights, and Commercial Use

Even if the first frame itself is an image you legally obtained, pay attention to how the final video can be used. Three points to watch:

First, portrait and personality rights. If the image shows an identifiable real person, even when the photo comes from a licensed stock library, using it to generate a video often still requires additional model authorization—especially for commercial promotion. The generated action may not match the person's actual behavior, creating a risk of misrepresentation.

Second, the model's terms of service. Different platforms have different rules about commercial rights to generated content. Some require a paid subscription to grant commercial use rights; others claim partial rights over generated content. Read the terms of service for the model you use before relying on it.

Third, possible training data contamination. Some models have been exposed to copyrighted material during training, and their outputs can occasionally carry recognizable styles or elements. If a generated result looks "like" a living artist or a well-known IP, the safest move is not to use it or to regenerate.

Summary

Image-to-video works best when you treat it as a two-step process. First, lock the frame with a strong static image, then direct a single clear moment of action with simple camera instructions. Keep the prompt short and specific, check your settings before submitting, and review the whole clip against your intention.

One-line Summary

Image-to-video anchors the frame with a still image and focuses the prompt on motion, making short video results more controllable.

Paragraph Summary

Image-to-video (I2V) is an input mode in which a generative video model receives a static first frame and then generates subsequent action based on a text prompt. Compared with pure text-to-video, it hands visual decisions to the image and motion decisions to the text, making results more controllable. This guide lays out the full workflow: first decide the use case—use text-to-video when the frame isn't set, and image-to-video when it is (product shots, character designs, key visuals); then prepare a first frame with a clear subject, room for motion, the correct aspect ratio, and clean details; decide on a single event and write in one sentence "what the viewer sees"; write a motion prompt covering only the subject's action, camera behavior, and details that must be preserved; check input mode, format, length, and cost before submitting; and finally review the whole video against your original intention. The article also lists four common problems (too much action, subject drift, camera wobble, barely any movement) and their fixes, and reminds you that the first frame needs clear rights and that portrait authorization deserves attention.