Seedance Image-to-Video Prompts That Keep Faces and Motion Stable

You upload a good photo, write a nice prompt, and two seconds into the clip the face belongs to someone else. Or the hands melt, or the product label turns into soup. With Seedance image-to-video this is rarely the model being weak. It is almost always the prompt asking for more than one photo can support.

This guide is a set of prompt rules and copy-ready prompts for Seedance 2.5, the ByteDance model the Dola app makes its videos with, that keep faces and motion stable. Everything here works in the Dola AI image to video tool, where your photo becomes the first frame of a 5 to 30-second clip with sound.

Why faces and motion drift

Three things cause most of the drift:

• Describing the photo. The model can already see it. When the prompt says "a beautiful woman with long brown hair", the model tries to match the words as well as the picture, and the face slides toward the words.
• Too many actions. Turn, wave, laugh and walk in five seconds means four large changes to the frame. Each one is a chance for the face to be redrawn.
• Small or soft faces. A face that fills 5% of a blurry photo gives the model very little to hold on to.

Fix those three and most clips stay stable.

The five-part prompt

Write every image-to-video prompt in this order:

1. The main action: one verb, one subject. "She turns toward the camera and smiles."
2. Secondary motion: small things that move around the subject. "A light breeze moves her hair, petals drift past."
3. The camera: one move. "Slow push-in."
4. The hold line: what must not change. "Keep her face and outfit exactly as in the photo."
5. Sound (optional): "Soft wind, distant birds."

The hold line does more than it looks. It tells Seedance which parts of the picture are fixed and which are allowed to move.

Prepare the photo first

• Crop to the final shape. A first-frame clip keeps the shape of your photo, so crop to 9:16 for TikTok and Reels, or 16:9 for YouTube, before you upload.
• Give the face room. Crop so the face is large in the frame. A head-and-shoulders crop holds identity far better than a full-body shot from across the room.
• Sharp beats pretty. A slightly plain but sharp photo beats a stylish, soft-focus one.
• Old prints: photograph them straight on, in even light, and crop away the border.

Eight prompts to copy

Each one follows the five-part order. Swap the subject words for yours.

1. Portrait, gentle

The person blinks, then slowly smiles and tilts their head slightly. Their hair moves a little in the breeze. Static camera, eye level. Keep the face, hair and clothes exactly as in the photo.

2. Portrait, turning to camera

She turns her head from the side toward the camera and holds eye contact. Background lights shimmer softly. Slow push-in. Keep her face and outfit exactly as in the photo.

3. Walking toward the lens

He walks slowly toward the camera, arms relaxed, city traffic blurred behind him. The camera moves backward to keep him in frame. Keep his face and jacket exactly as in the photo.

4. Pet

The cat looks up, ears forward, and slowly blinks. Its whiskers twitch. Static camera, low angle. Keep the fur colour and markings exactly as in the photo. Soft purring.

5. Product

Light sweeps slowly across the bottle from left to right, a few drops of water run down the glass. Slow orbit to the right. The bottle shape and label stay sharp and readable.

6. Old family photo

The people in the photo breathe and blink, one of them smiles softly. Very slow push-in. Keep the faces, clothes, film grain and colours exactly as in the original.

7. Illustrated character

The character raises a lantern and looks around, cloak moving in the wind, snow falling. Smooth tracking shot. Keep the art style, colours and character design exactly as in the image.

8. Couple

The two of them lean their heads together and laugh. Warm evening light flickers. Static camera. Keep both faces exactly as in the photo.

Notice what is missing: no description of hair colour, clothes, age or setting. The photo already says all of that.

Length: one beat per five seconds

Seedance 2.5 can run one take up to 30 seconds, but stability comes from pacing, not length. A good rule is one main action per five seconds. For a 15-second clip, write three beats in order: "First she looks up from the book. Then she closes it and smiles. Finally she stands and walks out of frame."

For anything longer than one take, make several clips and join them, starting each new clip from a still of where the last one ended. If what you are making is really a narrated story with the same character across many scenes, ToonBee runs Seedance 2.5 as well and has a story video mode that keeps the same characters in every scene, which saves stitching by hand.

Pick the right mode

• First frame: the default for one photo. The clip starts exactly on your picture.
• First & last frame: add a second photo and the clip moves from one to the other. Use it when the ending matters, such as a product turning to show its front, or a person standing up from a chair.
• Reference: attach several images and mention them in the prompt with @, for example "the jacket from @Image1". Use it when you want a character or object to appear in a new scene rather than animate an existing photo, and keep the hold line here too.

Restyle the still before you animate it

If you want a different look, such as a cartoon, a painting or a Halloween transformation, change the photo first and animate the result. Asking a video model to restyle and animate at the same time is the fastest way to lose the face. For Halloween, for example, an AI zombie filter that keeps your face recognisable turns a selfie into a zombie still in one of six styles, and that still can then be your first frame. Write the motion prompt for the new image, and keep the hold line pointed at the new style: "Keep the face and the zombie makeup exactly as in the image."

When the face still drifts

Symptom	Fix
Face changes when the head turns	Smaller turn, slower camera, add the hold line
Hands warp	Hands out of frame, or "hands stay still"
Product label melts	Closer crop, slower movement, "label stays sharp and readable"
Clip feels rushed	Fewer actions or a longer duration
Background warps	Static camera, or let only small things move
Two faces blend together	Crop tighter, one main action for both

Change one thing at a time between attempts so you know what fixed it. Test prompts at 480p, where each attempt costs fewer credits, and render the keeper at 720p or 1080p.

Same rules, other models

The motion-only prompt is not a Seedance trick. Most image-to-video models treat the photo as the first frame and the text as instructions for change, so the same five-part prompt carries over. Flow AI Studio's image-to-video tool gives the same advice for Google's Veo 3.1 (describe the movement, not the photo) and also supports a first and a last frame, so a prompt you tuned here is a fair starting point there. You can compare the video models available on Dolai on the AI models page.

Quick checklist

• Photo cropped to the final shape, face large and sharp.
• One main action per five seconds.
• One camera move.
• A hold line naming what must not change.
• No description of what is already in the photo.