Seedance 2.5 image-to-video prompts: direct the motion, not the picture

Text-to-video prompts describe a scene. Image-to-video prompts direct a shot. Your image has already decided the subject, the wardrobe, the light and the composition - so words that re-describe any of that are wasted, and words that contradict it cause drift. Every strong image-to-video prompt has three parts:

• What moves. Exactly one subject motion: "she turns toward the window and smiles", "steam rises from the cup", "the liquid pours in". One, not a choreography.
• What the camera does. One move - a slow push-in, a slow quarter orbit - or the legitimately great instruction "the camera holds steady".
• What stays fixed. Identity locks: "her face and the label stay unchanged", "the background stays still". These quiet the model's urge to improve your photo when nobody asked.

Fifteen to forty words is a complete image-to-video prompt. If yours is longer, it is probably re-describing the picture.

Portraits and characters are the most common starting point - a headshot, a character still, a photo of yourself. The motion should be small enough to be believable and large enough to feel alive:

Notice what is absent: no re-description of who she is, what she wears or where she is. The photo owns that. The prompt owns time, not space.

Product shots are the highest-value use of this mode, because the photo can be your actual product - accurate label, real materials, true color. Keep motion slow and let physics do the glamour:

For food, steam and pours are the two most reliable motions in the entire mode. For anything with printed text, keep the moves slow and add the "stays unchanged" lock every time.

Landscapes, city scenes and interiors animate beautifully, because almost anything can serve as believable motion - light, water, people, weather:

Illustrations, anime stills and paintings animate well because the style is baked into the image - the prompt only adds parallax and life:

Keep illustration motion gentler than photo motion - stylized art tolerates less movement before lines start to wobble. More templates and style blocks live in the anime prompt guide.

Why it fails: half the prompt re-describes what the photo already shows, then three actions are stacked on one camera instruction that says nothing. The model compromises on everything and delivers mush: a slightly different woman, doing three things badly.

Why it works: twenty-three words, one motion, one camera note, explicit locks. Every word directs instead of describing. When a take still drifts, tighten the locks and slow the motion before you touch anything else.

• Crop the photo first. The clip inherits the shape of the source image, so cut it to 9:16 or 1:1 before uploading, not after generating.
• Sharpest photo wins. The model cannot add detail the source lacks; a soft phone-zoom photo gives you a soft clip.
• Draft at 720p, finish at 1080p. Judge motion and drift on the draft, spend full resolution on the take you are keeping.
• Scale the motion words. "Slightly", "gently" and "slowly" are the most valuable words in this mode. Delete them when you want drama, but expect more drift in exchange.

One animated photo is a moment; a video is a sequence of them. The usual shape: pick the five photos that tell the story, animate each with a one-motion prompt, then cut them together with captions. Choosing photos that animate well is a skill of its own - sharp, well-lit, one clear subject, some breathing room - and the full image-to-video guide covers that side in depth.

When the shots are done, Seedance 2.5 Online keeps the sources, the takes and the timeline in one browser project, so assembly does not mean downloading everything and starting a second tool. The camera movement guide explains every move used in the templates above.