Crazy Frog AI Video Prompt: Copy-Ready Seedance 2.0 Templates

A prompt for the Crazy Frog AI trend has one job: describe a chase that reads in three seconds, while telling the model not to draw the motorcycle. Most failed attempts are not “bad prompts” in the literary sense. They are prompts that contradict themselves, leave the camera undefined, or ask for too many things at once.

This page gives four templates you can paste into the studio or another Seedance 2.0 workflow, then explains the lines that matter so you can adapt them without breaking the scene.

The four templates

1. The classic overpass

A real person crouches low on an invisible motorcycle, fists out on handlebars that are not there, racing down a concrete highway overpass at dusk. A police officer on a white hover-bike with flashing red lights closes in behind them. The rider glances back, leans into a curve, then bounces into a mid-air flip and lands back in the riding pose, speeding away. Low tracking camera, headlights streaking past, cinematic dusk light. No visible motorcycle under the rider. No text, no watermark.

Use this when you want the version that reads immediately as the trend. It is the most forgiving template because the overpass is visually simple.

2. The Halloween night

A person crouches on an invisible motorcycle and races down a suburban street on Halloween night, fists out on handlebars that are not there. Carved glowing jack-o’-lanterns line the road, string lights hang between trees, and a police officer on a white hover-bike with red lights follows close behind. The rider glances back, bounces into a playful mid-air flip and lands in the crouch, then speeds toward the fog. Warm porch light, cool blue dusk, low tracking camera, no visible motorcycle under the rider, no text, no watermark.

This one works with a single portrait and a vertical frame. The slower motion suits a softer, less aggressive joke.

3. The circuit board chase

A rider crouches on an invisible motorcycle on a glowing circuit-board road under a city overpass at night, fists out on handlebars that are not there. Green and cyan circuit traces light the road, and a police officer on a white hover-bike with red lights follows close behind. The rider glances back, flips mid-air and lands in the riding pose, then speeds into the distance. Low tracking camera, cinematic reflections, no visible motorcycle under the lead rider, no text, no watermark.

The second person is the point. Keep their role explicit (“chaser close behind”) so the model does not merge the two faces.

4. The desert run

A lone rider in a riding crouch speeds along an empty desert highway at golden hour, fists out on invisible handlebars, a dust trail behind them. Red rock mesas and a wide sky frame the road. The rider glances over one shoulder, bounces into a quick flip and lands still crouched, then shrinks into the distance. Warm sunlight, long shadows, side-tracking camera, no visible motorcycle, no text, no watermark.

This is the most “cinematic” of the four and the easiest to generate with one photo. The empty road gives the model fewer things to get wrong.

The lines that actually matter

“Invisible motorcycle” needs a negative instruction. Saying “riding an invisible motorcycle” is not enough; the model may interpret “motorcycle” as an object to draw. Add “no visible motorcycle, wheels, frame or handlebars under the rider” and repeat the idea in the action description. If you mention a bike in one line and forbid it in another, the model has to choose; keep the only bike in the scene under the chaser.

Define the camera. “Low tracking camera” is doing real work. Without a camera instruction, the model may cut between shots or drift into a static portrait. A tracking shot also hides hand and foot artifacts because the frame is moving.

Give each person a role. For two-photo renders, name the lead and the chaser in the prompt (“the person in the first reference photo is the lead rider; the second is the chaser”). If you do not, the model can swap them or blend their faces.

Keep the action to four beats. Launch, chase, flip, getaway. Adding a fifth beat (a wheelie, a conversation, a costume change) usually produces a cut or a confused pose. If you want a different action, replace a beat rather than adding one.

Say what not to render. “No text, no watermark, no visible motorcycle” is a useful tail. It is not a guarantee, but it reduces the odds of a stray logo or a bike frame.

How to adapt a template

Change the setting first. Replace “concrete overpass at dusk” with “rooftop parking garage at sunrise” or “seaside road at golden hour”. Keep the camera, the beats and the negative instruction. The setting is the part the model can vary without breaking the format.

Change the mood second. “Cinematic dusk light” can become “harsh midday sun” or “blue hour with street lamps”. Mood words affect color and contrast more than geometry.

Change the cast last. Adding a second person requires a second reference photo and a role sentence. Removing a person is easier: use a one-photo style and delete the chaser line.

What not to put in the prompt

• A celebrity name or a character name. The model may refuse or produce a lookalike you cannot use. Use a photo of yourself or an adult who agreed to appear.
• A minor. Do not generate a child in the chase. The site is for adults.
• “Make it go viral.” This is not a visual instruction. Describe the frame instead.
• A specific song lyric or a request to include the original recording. The model does not have the rights, and the download is silent. Add licensed sound in the app where you post.
• Contradictory vehicle instructions. “A motorcycle that is invisible but also visible in reflections” is a coin flip.

Using a prompt without a photo

The same templates work as text-to-video prompts. Remove the reference-photo sentence and describe the rider: “a person in a denim jacket and white sneakers”. The result is a scene with an invented face, useful for testing a setting or making a background plate. It is not a likeness of you.

A note on model behavior

Seedance 2.0 is a generative model, not a compositor. It does not follow a prompt like a recipe; it samples a plausible video. Two runs from the same prompt can differ in the face, the hands and the background. A clear photo and a short, consistent prompt reduce the variance. They do not remove it.

If a render fails, the credits are returned. If it succeeds but the result is not usable, change one variable: the photo, the setting or the length. Changing all three at once makes it impossible to learn what worked.

Try the built-in versions

The studio ships these four prompts as styles, with the framing, length and photo slots already set. You can start from a style and edit the prompt in place. If you want the full workflow, read the step-by-step guide; if you want to understand the meme first, read what the trend is.