Start from a strong still frame
The first frame determines most of the video. If the portrait has a clear face, a readable costume silhouette, and a simple background, the model has less work to do. If the still is already busy, the video will amplify the confusion.
Generate or edit the image first, then animate the version you actually like. Do not ask the video model to fix a costume that is already wrong.
Use one motion and one atmosphere
A good five-second prompt has one subject motion and one environmental motion. Head turn plus fog. Blink plus lantern flicker. Small step plus fluorescent flicker. More than that and the model starts inventing transitions.
Camera movement is a third layer. Keep it slow and simple: push in, pull out, or a short dolly. A dramatic orbit around a person is much harder to keep stable.
What to avoid in a Halloween video
Avoid dialogue, lip sync, complex hand gestures, dancing, running, and scene cuts. Those instructions ask the model to invent too much and usually produce warped faces or extra limbs.
Avoid adding creatures or a second person in the first pass. If you need a group scene, build it as a still image first and animate only the smallest movement.
Sizing the clip for social platforms
A vertical 9:16 clip works for stories and short-form video. A 16:9 clip works for a website hero or a YouTube pre-roll. Generate the still in the aspect ratio you need instead of cropping a wide image later.
Five seconds is enough for a loop. Trim the beginning and end if the model needs a moment to settle, then add music in your editor rather than asking the model to generate audio.
A five-second storyboard you can reuse
A reliable five-second clip has three beats. In the first second, hold the starting frame so the viewer can read the face and costume. In the next three seconds, add one small motion and one atmospheric change. In the final second, settle back into a clean pose that can loop or hold as an end card. This rhythm works for a portrait, a pet, and a product shot alike.
Write those beats into the prompt instead of asking for a full story. For example: hold the composition, turn the head slowly toward the camera, let fog drift across the lower frame, then settle. The model does not need to know what happens after five seconds. A clear ending frame also makes the clip easier to edit into a longer video with music and captions.
The cleanest result is usually the shortest prompt with the fewest moving parts.