Writing prompts for image-to-video models: what actually changes the output

Most prompt advice for AI video is written for text-to-video. When you start from a photo, the rules shift: the model already knows what is in the frame, so your prompt only needs to describe how it moves. Below are the prompt habits that tend to matter most with Seedance 2.5 based tools such as Dola AI.

1. Describe motion, not content

A prompt like "a woman in a red coat standing on a bridge" wastes tokens, because the photo already shows that. A better version:

slow push-in, coat moving slightly in the wind, river flowing, she stays still

Every clause is about motion, which leaves the model less room to reinvent the subject, so faces and clothing tend to stay consistent.

2. Pin what should not move

Without an explicit anchor, image-to-video models tend to animate everything a little: buildings breathe, text on signs melts. Adding one "stays still" clause usually reduces that:

camera orbits left, steam rises from the cup, table and background stay still

3. One camera move per clip

Combining "zoom in while panning right and tilting up" often produces jittery output. A single named move (push-in, pull-out, orbit, pan, handheld) is far more predictable. If you need two moves, generate two clips and cut them together.

4. Source image quality beats prompt length

The input often matters more than the text:

• crop so the subject fills at least a third of the frame
• avoid heavy compression artifacts, they turn into flicker
• avoid text overlays and screenshots, they rarely animate cleanly

5. Test at 5 seconds, then extend

Generation credits go a lot further if you iterate on a 5 second version until the motion looks right, then render the 10 second version once.

A simple checklist

If you want to try this workflow without installing anything, the Dola AI image to video page runs it in the browser and exports without a watermark. The same habits apply to other image-to-video tools as well, so they are not tool specific.