Crazy Frog AI With Friends: Make a Two-Person Chase

The one-photo version of the Crazy Frog AI trend is a solo joke: you, an invisible motorcycle, a generated cop. The two-photo version is a better joke, because the person chasing you is someone the group chat knows. It is also the version that most often goes wrong, because two faces give the model twice as many things to keep consistent.

This guide covers the two-person workflow: what the second slot actually does, which photo to use, how to write the role sentence, and what to do when the faces blend.

What the second slot does

The studio passes both images to the video model as reference inputs. The first photo is the lead rider — the person crouched on the invisible bike. The second photo is the chaser — the person the model puts on the hover-bike. If the second slot is empty, the model generates a cop that fits the scene; that is the one-photo workflow.

The chaser is not a face swap onto a fixed actor. The model generates the chaser’s close-up and the chase shots with the second reference as identity guidance. That means the chaser’s outfit, hair and approximate age come from the second photo. A friend in a bright red shirt will usually be recognisable; a friend in a dark photo with their face turned away will not.

Which styles support two people

Highway mecha is the two-photo style: 16:9 and 12 seconds by default. Circuit board, Desert run and Halloween night are one-photo styles; the second slot is hidden because those presets are built around a solitary rider.

You can change the aspect ratio and length after choosing a style, but the number of photo slots is part of the preset. If you want a vertical two-person clip, choose Highway mecha and switch the aspect ratio to 9:16. The prompt still describes two roles; the frame changes.

The second photo checklist

The same rules as the first photo, with one addition: the two photos should be easy to tell apart.

• One adult, alone in the frame. A couple or a group photo makes the model choose a face.
• Face fully visible. No sunglasses, no mask, no hand across the mouth.
• Even light. A dark photo next to a bright one makes the chaser look like a different person.
• A distinct outfit. A red shirt versus a denim jacket gives the model a visual anchor for each role.
• Recent and sharp. A blurry photo produces a blurry chase.
• Permission. The person must have agreed to appear, and must be an adult.

If the second photo is weak, leave the slot empty. A generated cop with a consistent look is better than a chaser who flickers between two faces.

Write the roles explicitly

The preset prompt already names the lead and the chaser, but if you edit the scene, keep the role sentence. A useful pattern:

The person in the first reference photo is the lead rider on the invisible motorcycle. The person in the second reference photo is the chaser on the white hover-bike. Keep both faces consistent with their reference photos. Do not swap their roles.

The last sentence matters. Without it, the model can put the chaser in the lead shot because it is the more visually prominent frame. The result is a clip where the wrong friend does the flip.

Keep the chase readable

A two-person chase has more going on, so the camera needs to stay simple. A low tracking shot that keeps both riders in frame is easier to generate than a sequence of close-ups. If you want a close-up of the chaser, ask for one beat only: “a brief close-up of the chaser, then back to the lead rider.” A prompt that asks for four close-ups and a wide shot in twelve seconds will produce cuts that do not match.

The four beats still apply: launch, chase, flip, getaway. The chaser is present in beats two and three; the lead rider owns the launch and the getaway. If you give both riders a flip, the model has to decide who lands first, and the timing falls apart.

When the faces blend

The most common two-person failure is a hybrid face: the model averages the two references into a person who looks like neither. The fix is usually the photos, not the prompt.

1. Make the photos more different. Different hair length, different clothing, different background. The model uses these as anchors.
2. Use brighter, sharper images. A low-contrast photo gives the model less identity signal.
3. Shorten the clip. Fewer frames mean fewer chances for the identity to drift.
4. Reduce the cast. If the second face is not essential, use a one-photo style and let the model generate the cop.

A second failure is role swapping: the chaser appears as the lead. Re-read the prompt and make the role sentence explicit. If it is already explicit, check that the first slot really contains the person you want as the lead.

Sharing the result

The finished clip is a single MP4 with both people. Download it and send it to the group chat, or use the email delivery field to send a private link to someone who does not have an account. The link expires after seven days.

If the clip is going on a public feed, get a second yes from the friend before posting. The uploader confirms permission, but the social decision is separate. A clip that is funny in a private chat can feel different when it is public.

A good first two-person run

Choose Highway mecha. Put the person who will do the flip in slot one and the friend in slot two. Use two photos with clearly different outfits. Keep the preset prompt, generate at 720p, and watch the first three seconds: if both faces are recognisable and the roles are correct, generate the 1080p version. If not, swap the photos before changing the prompt.

For a solo clip, read the one-photo workflow. For the prompt language behind the roles, read the prompt templates.