Image-to-Video Prompt Builder

Describe the motion you want — not the picture the model already sees

Image-to-video — 图生视频 — is a different craft from writing a still-image prompt. You already have a picture you like, and your job is to bring it to life without breaking what makes it good. The crucial mental shift is this: the model can already see your first frame, so re-describing the static composition in detail is mostly wasted effort. What it cannot see is the future — how the subject should move, where the camera should travel, how fast, and what must stay the same. This builder keeps that structure for you. You give a brief line on what is in the frame, then spend your words where they actually matter: the desired subject action, the camera movement, the motion strength, the duration, the consistency requirements, and a short negative list of artefacts to avoid. The result is the kind of prompt an experienced motion artist would assemble by hand, only built in seconds and portable across Jimeng, Kling and Hailuo.

The single most valuable instruction is the desired motion. "Slowly turns her head and smiles, hair drifting in a light breeze" tells the model exactly what to animate, far more reliably than a paragraph of mood adjectives. After the motion comes the camera: push in, pull out, pan, orbit, or a locked-off shot. Naming the camera move explicitly is far more controllable than leaving it to chance, and a locked-off camera paired with gentle subject motion is one of the most reliable starting points. Duration then sets the pacing — most clips run only a few seconds, so a single clear action usually beats trying to choreograph three. Treat each field as concrete and specific: instead of "make it dynamic", say "subtle, natural breathing motion; keep the face and hands stable".

Moderate amplitude and a steady camera are what keep the subject intact

The field that quietly saves the most footage is motion strength. Image-to-video models are sensitive to dramatic movement, and the moment you push the amplitude too high, faces stretch, hands grow extra fingers, and objects warp. Keeping the strength moderate — and reserving the bigger amplitudes for things that move naturally, like wind, water and smoke — is the cheapest way to avoid the most common failures. Pair that with a short negative prompt ("warping, distortion, flicker, extra limbs") and a clear consistency request ("keep the character's face, clothing and the first frame's lighting"), and the subject is far more likely to survive the clip looking like itself. None of this limits the model; it focuses it on a motion it can actually render cleanly.

Because the output is structured plain text, the same prompt is portable across every major Chinese video model, and the same principles carry to other image-to-video tools too. Write it in Chinese when you are working with Chinese models — the structure travels regardless of language. And because the whole tool runs locally in your browser, you can iterate freely: tweak the motion line, copy again, regenerate, and compare — without your text ever leaving your device or any model being called. Treat the first clip as a draft. If the subject drifts, dial the strength down; if it feels static, nudge one specific action up; if the camera fights the subject, lock the camera and let the subject move alone. Two or three rounds of that disciplined adjustment usually turn a warping, jittery first attempt into a clean, believable shot — and you keep a tidy, reusable prompt at the end.