A still is the first frame, not the whole video
Image-to-video does not interpolate a photo like a slideshow. It generates a new sequence of frames, using the still as the starting composition and the prompt as the motion brief. That is why a small, specific movement works and a vague "make it spooky" does not: the model needs to know what changes and what stays still.
The workflow on this site is deliberately split. You make the costume image first, check the face, then animate it. Doing both at once makes it hard to tell whether a bad result came from the image prompt or the motion prompt.
The motion grammar that works
A reliable five-second Halloween clip has four ingredients:
- One subject movement. A slow head turn toward the camera, a small head tilt, a blink, a tail wag.
- One environmental movement. Fog drifting, hair moving, a curtain shifting, a light flickering.
- One camera movement. A slow push-in, a gentle lateral drift, or a locked-off frame. Pick one.
- A consistency sentence. "Keep the face, costume, hairstyle and body proportions consistent with the reference."
Anything beyond that is usually too much for five seconds. Two people, a costume change, a camera orbit and a light cue in the same prompt will produce warping, not complexity.
Three tested motion presets
- Midnight Braids — AnimateA slow head turn, drifting hair and garden fog.
- Backrooms DriftA five-second liminal corridor push with flickering light.
- Pumpkin Pet — AnimateA tiny head tilt, ear twitch and a tail wag.
Midnight Braids — Animate is a portrait motion: the subject turns their head, loose hair moves, fog drifts, the camera pushes in. Backrooms Drift is the opposite kind of shot: an empty corridor, a forward walk, one flickering light panel. Pumpkin Pet — Animate is the smallest motion of the three: a head tilt, an ear twitch and two tail wags. Small motions are not a compromise; they are the highest-success-rate way to animate an image.
Portrait clips: keep the face stable
The face is the first thing a viewer checks and the first thing a video model can break. Four habits reduce drift:
- Start with a sharp still. A soft or low-resolution source gives the model less information to preserve.
- Keep the head turn slow. "Slowly turns their head toward the camera" is safer than "spins around".
- Avoid occlusions. Hair across the face, a hand on the cheek or a mask edge gives the model more to reconstruct.
- Lock the lighting. Say "consistent lighting" and avoid describing new light sources unless the light itself is the effect.
If the face still warps, cut the clip to three seconds and remove the camera movement. A locked-off frame with one small movement is the most stable result.
Corridor and scene clips: keep the geometry
For a Backrooms-style shot, the failure mode is different: walls bend, doorways move, and the corridor changes length. The prompt should protect the geometry explicitly: "keep the wall and doorway positions stable, no new objects, no cuts". One flickering light and a slow forward move are enough. You do not need a monster; the empty corridor is the effect.
The still for this shot can be text-to-image. Generate the corridor with the AI Halloween video generator's text-to-image sibling, the AI image generator, then animate it. A corridor with no people is also easier to generate than a convincing crowd.
Model choice and cost
The workbench offers several image-to-video families, including Kling, Sora and Veo lanes through the configured providers. They differ in motion quality, duration limits, audio support and price. A practical first pass is a fast, lower-cost lane; move to a higher-quality lane only when the composition and the motion are already right.
Video is billed per second, so a ten-second clip costs more than a five-second clip on the same lane. The workbench shows the exact credit cost for the selected model and duration before you submit. The pricing page lists the packs and plans.
Do not pay for a long clip to fix a bad still. Fix the still first. The video model can animate a good image; it cannot rescue a costume that does not read.
Finishing the clip
- Download the clip before you close the tab. Generated task URLs are temporary.
- Add sound in your editor. A low room tone, a distant hum or a single music cue does more for the Halloween mood than a more complex prompt.
- Keep a consistent grade. If you are making several clips, apply the same contrast and color treatment so they cut together.
- Disclose synthetic media where the platform requires it.
For the image side, the Halloween hub collects the costume, pet, couple and makeup pages. For the prompt text behind each motion preset, the Halloween AI prompts page has the full copy-ready versions.