Best Photos for Crazy Frog AI: A Portrait Checklist
Crazy Frog AI Video TeamThe photo decides most of the result. Here is the checklist for a rider portrait, the failure modes to avoid, and how to fix a weak source image.
只上传你自己或已同意出镜的人。不要上传名人和未成年人。
Keep the invisible-motorcycle idea in the prompt. Describe the rider, the road and the chase.
720p · 12 秒消耗 10 积分；1080p · 12 秒消耗 30 积分。更长视频按比例计算并向上取整。
本次消耗 30 积分 · Pro 月付 $39.90 约 6 条/月，年付 $319 约 80 条/年。
A Crazy Frog AI clip starts as a still photo. The model reads that photo for identity, hair, outfit and lighting, then generates the chase around it. No prompt repairs a photo that was never usable. Choosing a better source image is the highest-leverage thing you can do, and it takes about a minute.
This is the checklist we would give a friend: what a good portrait looks like, what the hard limits are, and the specific failure modes that produce strange results.
The short checklist
A photo that renders well is:
- One adult, alone in the frame.
- Face fully visible: eyes, nose and mouth unobstructed.
- Front-facing, looking toward the camera.
- Evenly lit: daylight from a window, open shade, or a soft indoor light.
- Sharp at full size on your phone.
- Unfiltered: no heavy beauty mode, no smoothing, no stickers.
- Waist-up or full body, so the outfit is visible.
- A JPG, PNG or WebP file, up to 10 MB.
That is the whole game. The rest of this article explains why each line matters.
Why the photo decides the result
When you upload a photo, the model is not animating a cut-out face. It reconstructs a person: head angle, shoulder width, hair volume, how the clothes drape, where the light falls. The prompt supplies the chase and the camera; the photo supplies the identity.
Everything the model cannot see, it guesses. A photo that stops at the chin forces it to invent a body. A face half in shadow forces it to invent the geometry. A hat pulled low forces it to invent a forehead. Guessing is where the uncanny results come from: a face that drifts, a jaw that slides, a rider whose head does not turn with the body.
The fix is rarely a cleverer prompt. The fix is a photo with less ambiguity.
Face and light
Front-facing beats a profile. The model needs both eyes and the full mouth to keep the face stable when the rider looks back. A three-quarter turn can work; a full profile usually does not.
Even light beats dramatic light. A single hard lamp creates a bright side and a dark side; the model has to invent the dark half. Window light or open shade gives it the whole face.
No sunglasses, masks or hands. Anything covering the face removes identity information. A clear face is more important than a cool pose.
No heavy filter. Beauty mode removes skin texture before the model sees it. If the smoothing is visible at full size, it is too much.
Body and outfit
The chase is a full-body pose, so the model needs to see more than a headshot. Waist-up is the minimum; knees-up or full body is better. The outfit matters because it is copied: a bright jacket or a distinctive pattern gives the viewer a way to recognise the rider across shots.
Avoid a seated pose, a heavy coat that hides the silhouette, or a photo cropped at the waist if you want a full-body rider. The model will invent legs, and invented legs are where proportion errors show up.
The photo failures that actually happen
The tiny face in a group photo. The model picks one face, often the largest, and ignores the rest. The result is a rider who may not be you. Crop to one person before uploading.
The dark selfie. A face lit only by a screen loses color and detail. The model produces a soft, shifting likeness. Take one more photo near a window.
The old photo. A photo from five years ago is fine if it still looks like you; a photo with a different hairstyle, weight or glasses will produce a rider who looks like a past version of you. That can be the joke, but decide on purpose.
The screenshot. Screenshots are compressed and may include UI. Use the original image.
The celebrity. Do not upload a public figure. The site is for you and adults who agreed to appear.
The second photo
For a two-person chase, the second photo follows the same rules, plus one: make the two people easy to tell apart. Different hair, different clothing, different background. If both photos look similar, the model can blend the faces or swap the roles.
If the second photo is weak, leave the slot empty and let the model generate a cop. A consistent generated chaser is better than a bad likeness.
How to test a photo cheaply
Generate the shortest, lowest-resolution version the studio allows. The cost badge updates before you press Generate. A 720p test answers the only question that matters: does the face hold in motion? If it does, generate the 1080p version. If it does not, change the photo before changing the prompt.
Do not judge the photo by how it looks as a still. A photo can be beautiful and still be a bad reference: too much shadow, too much filter, too little face. The test is the render.
A note on permission
Only upload a photo of yourself or an adult who has agreed to appear. Do not upload a minor, a celebrity or a public figure. The uploader confirms this before generating, and moderation may reject an input that looks like a public figure. If someone asks you to remove a clip, remove it; the terms make that the right response.
Related guides
A two-minute photo routine
If you do not have a usable portrait, take one now. Stand near a window with the light on your face, not behind you. Hold the phone at eye level, not below the chin. Take three frames: a neutral expression, a slight smile, and one with your shoulders visible. Pick the sharpest one, not the prettiest. That frame is the reference.
The routine is deliberately boring. The goal is not a great photograph; it is a photograph with enough information for the model. A neutral expression and even light give it the most to work with. Save the dramatic lighting for the clip, where the preset already supplies it.