Every loop on this site started as a written prompt. Nothing was bought as stock footage. Below is how to write one, and the actual prompts we use — copy them, swap the subject, generate something that is not on this site yet.
The order matters. Deciding the shape first is what stops you generating something beautiful that is invisible from the street.
Before any prompt: what will the street actually see? A crowd, one creature, a swarm crossing, or a lit object. Single silhouettes read through a curtain far better than busy scenes — a crowd turns to mush at ten metres.
Ask Claude or ChatGPT for a text-to-video prompt and give it the six ingredients below. Or take one of the examples on this page and swap the subject. The six ingredients are what make it projectable; skip them and you get a video, not a window.
Paste the prompt into a text-to-video tool and generate. Six to eight seconds is plenty, and 16:9 is the framing you want. You are looping it all night, so motion quality matters far more than length.
Play it on repeat. If the first and last frames do not match you get a visible jump every eight seconds, which is the single most obvious sign of an amateur projection. Regenerate rather than trying to fix it in an editor.
Put it through the projector onto your actual curtain after dark, and look from the street. Things that look great on a monitor — fine detail, mid-tones, small text — are often invisible once projected.
Every projectable prompt has these. Leave one out and you get a video rather than a window.
One clear thing. A hunched figure, a galleon, a swarm of bats.
The darkness is what the projector does not light up. This is the whole trick.
Bold shapes, no mid-tones. Detail is lost through fabric.
Drifting, swaying, crossing. Fast or complex motion blurs on a curtain.
Edges of a window get clipped by the frame and the curtain.
Rendered text turns to nonsense and looks like a mistake.
Write a text-to-video prompt for a Halloween window projection. Subject: [WHAT IT IS, e.g. a skeletal pirate crew on a galleon] Style: pure black background, high-contrast luminous silhouette, bold shapes Motion: [SIMPLE MOTION, e.g. drifting slowly sideways through low fog] Framing: centred, wide 16:9, nothing near the edges Exclude: no text, no watermarks, no borders, no camera shake Loop: the first and last frames must match so it repeats seamlessly Give me the finished prompt only, ready to paste into a video tool.
Six working prompts covering the shots each build needs. Copy one, change the subject, and you have a new theme.
The photorealistic shot of a house at night with a glowing window. Use this as your build's hero/street clip.
Night-time street view of a suburban house with a large window glowing deep orange. The silhouette of a skeletal crew on a galleon drifts across the window surface behind a pale sheer curtain. Photorealistic, cinematic, Halloween night atmosphere, handheld camera feel, low fog on the lawn, carved pumpkins, warm porch light.
Explains the method by showing it. Same prompt works for every build because the rig never changes.
Inside a dark room at night, a mini projector on a shelf throws a bright beam at a white sheer curtain covering a large window. Dust motes in the beam, the fabric glowing softly where the image lands. Photorealistic, moody, single practical light source, shallow depth of field.
The workhorse. Black background, bold silhouette, seamless. This is the only clip a visitor strictly needs.
Decorative Halloween projection animation on a black background: a skeletal pirate crew on a galleon drifting slowly sideways through low fog, tattered sails catching an unseen light, all in black silhouette. High-contrast luminous tones, bold silhouette shapes, clean composition centred in frame, no text, no borders, seamless looping motion.
Best for tight windows and close viewing. One subject reads through a curtain far better than a crowd.
Decorative Halloween projection animation on a black background: a hunched figure in a wide-brimmed hat stirring a bubbling cauldron, green vapour rising over the rim, flickering flames underneath, all in black silhouette. High-contrast luminous tones, bold silhouette shapes, clean composition centred in frame, no text, no borders, seamless looping motion.
Use when you want motion across a whole window. Keep individual shapes simple or they turn to mush.
Decorative Halloween projection animation on a black background: a stream of bats crossing the frame from left to right in loose waves, wings spread wide, the shapes growing larger as they pass nearest the viewer, all in black silhouette. High-contrast luminous tones, bold silhouette shapes, clean composition centred in frame, no text, no borders, seamless looping motion.
Bright single objects hold up on a lit street much better than dark scenes do.
Decorative Halloween projection animation on a black background: a weathered iron lantern swinging gently, its flame flickering and casting irregular light, moths circling the glass, all in black silhouette with a luminous core. High-contrast luminous tones, bold silhouette shapes, clean composition centred in frame, no text, no borders, seamless looping motion.
The prompts above work in any text-to-video tool. Here is what we use, and why.
One interface and one credit balance across several text-to-video models. That matters for this job specifically: silhouette motion is the hard part, and different models handle it differently, so being able to run the same prompt twice and keep the better result is worth more than any single model's quality.
Credits rather than a subscription, so a small top-up is enough to try it. Check the current models and rates on their site.
We earn a commission if you sign up through our link.
Loop the file fullscreen from a laptop or media player, project it at the back of a white sheer curtain from inside the room, and turn every other light off. Check it from the street before the night you need it.
The gear that makes this work — projector, curtain, fog — is on every build page, and the full setups are in the guides.
Cost, looping, why the background is black, and what to do when the output looks wrong.
We generate with KIE.ai, which runs several text-to-video models behind one interface and one credit balance — that means you can try the same prompt on a couple of models and keep whichever handles silhouette motion better. Our affiliate link is below; it costs you nothing extra and we earn a commission if you sign up.
Each clip costs a small number of credits rather than a subscription. A full three-clip build for one theme runs in the low hundreds of credits, so a whole month of nightly builds costs less than a single stock-footage licence. You can start with a small top-up and see how far it goes before committing.
The prompts on this page are written to be model-agnostic — they describe the shot, not a particular engine, so they work in any text-to-video tool. We use KIE because it gives access to several models from one balance, which makes comparing outputs cheap, but nothing here locks you to it.
Because a projector cannot project black. Black is simply the absence of light, so a black background stays dark on the curtain and only your silhouette glows. That is why these read as glowing shapes through fabric instead of as a flat rectangle of video.
Two reasons. AI video renders text as garbled shapes almost every time, and even when it works, text on a window is backwards from the street unless you mirror it. If you want words in the display, cut them as physical letters or generate a mirrored version deliberately.
Say so in the prompt — "the first and last frames must match so it repeats seamlessly" — and then verify by playing it on repeat before you project it. Generative tools do not guarantee this, so treat a visible jump as a reason to regenerate, not something to live with.
The prompts are free to use and adapt however you like, including commercially. What the underlying tool permits you to do with the output is its own licence, so check that if you are generating for a paid event or a client.
Regenerate before rewriting the prompt. Text-to-video output varies a lot between runs from an identical prompt, so one bad result is usually luck rather than a bad prompt. If five attempts all fail the same way, the prompt is asking for something the model handles badly — most often a busy scene or complex motion.