Tutorial · free route included

How to Make a Crazy Frog AI Video

There are two honest routes to this clip. Route A is free-ish and yours: a general AI video model plus the prompt below. Route B is ours: one photo in, one finished 14-second chase out, $9.90, no prompt, no editing. This page shows both, so you can pick with your evening intact.

Route A — DIY in a general AI video model (free steps)

This is how the original viral clip was made, and how most of the trend still operates. Budget an evening the first time; twenty minutes once you have the prompt saved.

Step 1 — Pick your model

Any current video model with image-to-video (a reference photo) can do it: Higgsfield, Google Veo, Kling, ByteDance Seedance, and the rest. You need two capabilities: reference-photo conditioning (so the rider keeps your face) and vertical 9:16 output. Free tiers exist on most of them — expect watermarks or queueing in exchange.

Step 2 — Prepare the photo

• Front-facing, face clearly visible, no sunglasses, good light.
• Upper body in frame — the model renders a better squat with more of you to work with.
• Yourself or someone who agrees to be in it. Not celebrities, not public figures — see our terms for the reasoning, which applies everywhere, not just here.

Step 3 — The prompt (copy-paste, then swap the scene line)

The whole trend lives or dies on one detail: the moped is invisible. Say it explicitly, in caps if your model respects emphasis, or you will get a perfectly visible scooter and a perfectly normal video. Paste this, change the scene line to wherever you want the chase:

Subject: the person in the attached reference photo — same face, same hairstyle, same outfit.

Scene: [pick one, e.g. "an empty city overpass at golden hour" / "a rain-slick Tokyo street at night" / "a desert highway"].

Action, six shots in one continuous 14-second vertical (9:16) video:
1. Close-up of a motorcycle cop on a white hover-bike, red and blue lights strobing, chasing the subject.
2. The subject riding an INVISIBLE moped at speed — squatting low, hands gripping unseen handlebars, body leaning into the ride. The moped is not visible; at most a single wheel shows.
3. The subject glances back over one shoulder and does a comic jump of fright.
4. On the beat, the subject launches into a full mid-air backflip against the sky.
5. Fast getaway: the subject rockets past the camera; the hover-bike falls behind and shrinks.
6. Final wide shot: the subject alone on the empty road, a tiny dot pulling away. The cop is gone.

Style: cinematic action-comedy, motion blur, engine-shake camera in the chase shots, epic slow shutter on the backflip, sound of engine roar and police siren. No visible moped body — that is the joke.

Generate 14 seconds at 9:16. If the model caps out at 10 or 12 seconds, generate the chase and the backflip as two clips — the storyboard above splits cleanly after shot 3.

Step 4 — Fix what breaks

Every failure mode has a standard patch:

• The bike is visible. Your model over-helped. Strengthen the negation: "no visible motorcycle body, rider floats in riding pose".
• The face drifts. Re-anchor with the reference photo per generation, not per session, and keep the face out of fast motion in shot 1 (start on the cop, like the storyboard does).
• No backflip. It is the hardest physics in the sequence. Split it into its own generation: "the subject launches from a squat into one full mid-air backflip, slow shutter, sky behind".
• Wrong aspect ratio. Set 9:16 before generating. Cropping a 16:9 chase afterwards beheads your cop.

Step 5 — Cut and post

If you generated multiple clips, cut them in any editor (CapCut is free) in storyboard order, hard cuts on the beat. Add music that you have the rights to — the trend's signature track is copyrighted, and here is the full answer on that. Engine roar plus siren, which most models generate natively, is genuinely enough.

Route B — one photo, one click, done

The generator on this site runs the same storyboard as a fixed pipeline: the six frames are drawn from your photo, then the video is filmed from those frames. You pick the background — the original overpass, or one of 16 more — and download the finished MP4: 14 seconds, 9:16, 720p, no watermark, native engine and siren sound. $9.90 per video, usually 4 to 6 minutes, auto-refunded if the render fails.

In other words: Route A costs an evening and maybe a subscription; Route B costs $9.90 and four to six minutes. The pricing page has the pack math if you are making these for a group chat.

Which model is best for this, honestly

As of October 2026: the clips coming out of ByteDance's Seedance 2.0 class of models have the best face-keeping and the cleanest backflips; Google's Veo generation leads on physics realism if you want the chase to look expensive; Higgsfield is the most trend-native — it ships effect presets built for exactly this kind of meme (see our comparison). This site's own pipeline draws its storyboard frames with Google's Gemini image model and films the video with Seedance 2.0.

One legality note before you post

Keep the rider to consenting humans, skip the original soundtrack unless you licensed it, and label it what it is — an AI clip. Every platform's 2026 rules reward the label and quietly bury the ones that hide it.