Image-to-video prompts work best when they describe the camera, not the picture. The picture already exists in frame 1. Your prompt should tell the model how the camera and the subject move, and what must stay the same.
This guide gives a five-line pattern you can use in any app, the wording for six common camera moves, and fixes for the problems you are most likely to see. Prompt Lab builds this pattern for you.
The five-line pattern
Subject: [who and where, one sentence]. Fictional adult, clearly over 25. Frame 1: [shot size and angle of the start image]. Camera and motion: [one camera move], [subtle | medium | strong] intensity. Format: [9:16 | 16:9 | 1:1], 24 fps, 5 seconds. Consistency: keep face, skin texture and lighting identical to frame 1. No text, no watermark.
- Subject repeats what is in the image in plain words. It anchors the model when the camera moves away from details.
- Frame 1 names the shot size, such as close-up or full body. A move that does not fit the shot, like an orbit on a tight close-up, often breaks.
- Camera and motion holds one move. Two moves in five seconds usually produce a muddled clip.
- Format should match what the model supports. For example, the Wan 2.2 TI2V-5B model card lists 720p at 24 fps.
- Consistency tells the model what must not change. Most identity drift starts here.
Wording for six camera moves
| Move | Prompt wording | Best shot |
|---|---|---|
| Push-in | slow dolly push-in toward the subject, shallow depth of field | Medium close-up |
| Orbit | smooth 180-degree orbit around the subject at chest height | Three-quarter or full upper body |
| Wind | gentle wind moving hair and fabric while the subject holds the pose | Medium shot with loose hair or fabric |
| Loop | rhythmic sway with feet planted, seamless loop | Full body on a plain background |
| Handheld | subtle handheld drift with natural micro-shake, 35 mm lens look | Medium shot |
| Portrait | identity locked, slow blink and a head turn under 15 degrees | Tight portrait with even light |
Worked examples by shot type
Copy any of the blocks below into your image-to-video tool or adjust the subject line in Prompt Lab. Keep the five lines in the same order so you can spot which line caused a change when you compare two outputs.
1. Tight portrait with subtle breathing and eye movement
Subject: Woman in her early thirties in a dimly lit studio, warm key light from the left. Fictional adult, clearly over 25. Frame 1: Tight close-up from the collarbones up, eye-level camera, 85 mm portrait framing. Camera and motion: locked tripod frame, slow blink and a head turn under 15 degrees, subtle intensity. Format: 9:16, 24 fps, 5 seconds. Consistency: keep face, skin texture, eye color and lighting identical to frame 1. No text, no watermark.
2. Medium shot with wind and fabric motion
Subject: Man in his thirties standing on a rooftop terrace at dusk, open linen shirt, city lights behind him. Fictional adult, clearly over 25. Frame 1: Medium shot from the waist up, slight low angle, soft rim light. Camera and motion: gentle wind moving hair and fabric while the subject holds the pose, slow dolly push-in, subtle intensity. Format: 16:9, 24 fps, 5 seconds. Consistency: keep face, jawline, shirt collar and background lights identical to frame 1. No text, no watermark.
3. Three-quarter shot with camera orbit
Subject: Woman in her late twenties in a velvet evening dress beside a marble column, warm indoor lamps. Fictional adult, clearly over 25. Frame 1: Three-quarter shot from mid-thigh up, clear space on both sides of the subject. Camera and motion: smooth 180-degree orbit around the subject at chest height, medium intensity. Format: 9:16, 24 fps, 5 seconds. Consistency: keep face, dress silhouette, hands at sides and room lighting identical to frame 1. No text, no watermark.
How to prepare frame 1 before you animate
Half of the defects people blame on the video model come from the starting image. Before you spend credits or GPU time, check four things in frame 1:
- Crop to the target aspect ratio first. Feeding a square image into a 9:16 or 16:9 video run forces the tool to invent new background or stretch the frame. Crop or outpaint the still image to match your video ratio before you animate it.
- Give the face enough pixels and even light. When a face occupies only a small patch of a wide shot, the video model has very little detail to anchor identity across 120 frames. Heavy shadow across one eye often turns into asymmetry as soon as the head turns.
- Tuck hands or rest them on a surface. Fingers waving in mid-air are the first thing to split or gain extra joints during motion. A hand resting on a hip, railing, or lap stays stable much longer.
- Avoid mid-blink or extreme neck angles in the still. If frame 1 catches the subject mid-squint or looking sharply over one shoulder, the model has to guess what the neutral face looks like as soon as the pose relaxes.
Common problems and what to change
The face changes during the clip
Lower the intensity, choose a smaller move such as a push-in, and keep the consistency line. Start from a frame where the face is large, sharp and evenly lit. Wide shots give the model fewer pixels of face to hold on to.
The camera move is ignored
Put the move first in the motion line and remove other actions. Words like “dynamic” or “cinematic” on their own do not describe a move; name the move.
Hands, limbs or the background melt
Ask for less motion in those areas, for example “hands stay still, background stays fixed”. Keep hands out of frame 1 when you can.
The clip does not loop
Say “seamless loop” and ask for a movement that returns to its start, such as a sway. Moves that travel, like a push-in, cannot loop cleanly.
Text or a watermark appears
Keep “no text, no watermark” in the last line. Some apps add their own watermark on lower plans; the prompt cannot remove that.
Why five seconds beats ten seconds for identity lock
Drift compounds over time: frame 60 is predicted from frames that have already drifted slightly from frame 1. When a tool offers both 5-second and 10-second clips, a 5-second generation with subtle or medium intensity holds facial features more reliably than a single 10-second pass with strong motion.
Keep it to fictional adults
Describe invented characters and say they are adults. Do not use a real person’s name or photo. Making intimate images of real people without consent is a crime in many places, and the apps we profile ban it in their own terms. Read our consent and safety policy.
Where to use the prompt
Cloud apps differ in which of these lines they respect and in what they allow. Our comparison of image-to-video apps lists what each one documents. To run open models on your own GPU, start with our Wan 2.2 ComfyUI setup guide.
Check current plans on Promptchan
Documents both text-to-video and image-to-video ('Animate') on its Pro plan.
This pattern is our editorial guidance from writing prompts for these tools. Model-specific limits, such as resolution and frame rate, come from the linked official pages, checked on October 31, 2026.