Image-to-Video Prompt Guide: Animate a Still Without Losing It

Image-to-video generation starts with a contract: the source image defines what the first frame looks like, and the prompt explains what should change over time. Weak results often come from breaking that contract. The prompt redescribes the image as if it were text-to-video, asks for a different composition, or demands motion that the still does not contain enough visual information to support.

A better workflow treats the image as production material. First check whether it is a usable keyframe. Then prompt motion, camera, and sound while leaving established identity and composition alone.

Google describes image-to-video as using an existing image as the first frame in its official Veo image-to-video documentation. That framing is useful beyond one model: your input is not merely a loose inspiration image. It is the visual starting state of the shot.

Decide what the image controls

A source image can anchor several things:

• subject identity and appearance;
• product shape, color, and label placement;
• clothing and accessories;
• environment and lighting;
• camera angle and composition;
• visual medium or texture.

List the non-negotiable anchors before writing a prompt. For a product image they may be bottle geometry, cap color, label typography, and liquid color. For a character they may be face, hairstyle, jacket, and proportions. The prompt should not casually overwrite these facts.

Instead of:

A luxurious blue perfume bottle on black marble, cinematic studio lighting, centered composition.

write:

The bottle remains unchanged. A narrow highlight travels slowly across the glass while faint mist drifts behind it. Locked-off close-up.

The source image already communicates the first sentence’s scene. The second prompt spends its limited attention on motion.

Prepare a source frame with room to move

Image quality is not only resolution. The frame needs spatial logic for the requested action.

Check:

1. Subject completeness. If a person’s hand is cropped at the wrist, asking that hand to reach across the frame forces the model to invent anatomy outside the image.
2. Movement space. A runner placed against the right edge has nowhere to run right. Leave negative space in the direction of travel.
3. Occlusion. Hidden legs, handles, wheels, or product edges may reappear inconsistently when the camera moves.
4. Background depth. An orbit or dolly needs foreground, subject, and background layers. A flat graphic may support a zoom but not convincing parallax.
5. Readable detail. Small logos and dense text are fragile across frames. Use a clean label and plan to add critical copy in editing.
6. Aspect ratio. Compose the source for the final output. Cropping a wide keyframe into vertical video after generation may remove the very motion you asked for.

Do not upscale a poor source and assume the problem is solved. Sharper compression artifacts are still artifacts. Start from a clean image with intentional framing.

Prompt the change, not the still

An image-to-video prompt should answer three questions:

• What moves?
• How does it move?
• What does the camera do while it moves?

For example:

The leaves sway gently in a crosswind and two small ripples travel across the pond. The camera makes a very slow push toward the cabin. Morning light remains consistent.

This separates environmental motion from camera motion and protects the established light.

For a portrait:

She takes one calm breath, turns her eyes toward the window, and gives a slight natural smile. Static medium close-up; facial features, hairstyle, and clothing remain unchanged.

For a product:

The turntable rotates the shoe 30 degrees clockwise at constant speed. Locked camera, clean white background, sole shape and branding remain unchanged.

Keep the action proportionate to the clip. A four-second image-to-video shot can support a glance, small turn, breeze, steam plume, or modest camera move. It cannot convincingly cover an entire unboxing sequence.

Choose internal motion or camera motion first

The source frame is easiest to preserve when only one kind of movement dominates.

Internal motion changes elements inside a stable frame:

• blinking, breathing, hair, cloth, steam, water, reflections;
• a hand making one small gesture;
• a product rotating on a fixed base.

Camera motion changes the viewpoint:

• slow dolly in;
• short lateral slide;
• limited orbit;
• tilt from one framed element to another.

Begin with internal motion and a locked camera when identity is critical. Add camera movement after the subject is stable. For a camera vocabulary and debugging sequence, use the camera movement prompt guide.

An orbit is especially demanding because it reveals parts of the scene absent from the source. Limit it to a short arc and use a subject whose unseen sides are predictable. A front-facing package with detailed printing on every panel is a poor candidate for an invented 360-degree spin.

Preserve products and text deliberately

Generative video may reinterpret small symbols and typography from frame to frame. You can reduce, but not eliminate, this risk:

• make the product large in the source frame;
• use a slow, limited motion;
• avoid heavy motion blur across the label;
• keep the camera distance stable;
• state that geometry, label layout, and colors remain unchanged;
• inspect individual frames, not only playback;
• composite mission-critical text or logos in post-production.

Do not promise exact label preservation to a client until the actual output has passed frame-level review. If regulatory text, dosage, pricing, or a legal mark must remain perfectly readable, generative rendering should not be the final typography layer.

Animate people without overloading anatomy

For portraits, subtle motion is usually more credible than choreography. Useful instructions include:

• natural blink once;
• slight head turn;
• one breath visible in shoulders;
• small shift of gaze;
• hair moving gently in wind;
• one hand entering frame for a simple action.

Avoid asking a cropped portrait to stand up, turn around, and walk away. The model lacks a defined body, back view, and environment beyond the frame. Generate or photograph a wider source image first.

Describe expression through visible changes rather than emotion labels:

Her eyebrows relax and the corners of her mouth lift slightly.

is more concrete than:

She becomes inspired and hopeful.

If a real person appears in the source, make sure you have permission to animate the likeness and do not create deceptive or harmful portrayals. Model access does not replace consent, publicity rights, or platform rules.

Add atmosphere in layers

Atmospheric motion makes a still feel alive without forcing the primary subject to transform:

• drifting dust caught in backlight;
• moving reflections;
• steam or condensation;
• curtains lifting in a breeze;
• shallow water ripples;
• passing soft shadows;
• distant pedestrians or traffic.

Choose one or two. A frame containing rain, smoke, sparks, birds, crowds, blowing papers, flashing signs, and a moving camera has too many independent systems. Background motion can also cross the subject and damage its silhouette.

A layered prompt might read:

The ceramic cup remains still. Steam curls upward in two thin strands while a soft window reflection moves across the table. Slow 5 percent push-in; no change to the cup shape or printed mark.

Use negative prompts narrowly

If the model supports a negative field, list visible failure classes rather than commands:

warped label, changing product shape, duplicated handle, camera shake, abrupt zoom

Do not paste a universal hundred-term negative list. It can conflict with the desired motion and makes diagnosis impossible. Generate once, identify the repeated defect, then add only the relevant exclusion.

Debug image-to-video drift

The subject changes immediately

Use a cleaner, larger source subject. Remove a long visual restatement from the prompt. Start with locked camera and minimal movement.

The background melts during a dolly

The source may not define enough depth. Reduce the move, switch to a zoom, or prepare a frame with stronger foreground/background separation.

A face changes midway

Reduce head rotation, camera angle change, and expression intensity. Use a higher-quality facial reference and a shorter clip.

A product grows extra parts

Keep all important edges visible in the source. Avoid rotations that expose unknown surfaces. State one small rotational range and fixed geometry.

Motion is too weak

Make the action observable and directional: “two steps toward camera” is stronger than “moves naturally.” Remove style language that may compete with motion.

Motion is chaotic

Use one action, one camera instruction, and a constant tempo. Remove words such as dynamic, dramatic, explosive, and fast unless they are truly required.

A production checklist

Before generating:

• final aspect ratio chosen;
• clean source without accidental crops;
• negative space in the direction of action;
• critical identity anchors listed;
• one subject action selected;
• one camera behavior selected;
• output duration fits the motion.

After generating:

• scrub frame by frame around hands, faces, edges, labels, and reflections;
• compare first and last frames for shape and color drift;
• verify the camera path does not cross objects;
• check loops for a visible jump;
• add critical text, logos, captions, and audio in a deterministic editor;
• retain the source and prompt with the approved clip for reproducibility.

The HyperFrames workspace supports both text-to-video and image-to-video scenes across available models. Model choice can improve a difficult shot, but the largest quality gain often comes earlier: choose a source image that contains the information the requested movement needs, then prompt only the change.