How to make an AI rap video

1. Pick the right photo

The photo is the single biggest factor in the result. What the model needs is a clear view of your face: one person, facing the camera, head and shoulders in frame, even light with no strong shadow across the eyes or nose.

What breaks it is anything that hides the face or changes its geometry. Sunglasses remove the eyes, which are where a viewer decides a face is real. Heavy beauty filters smooth away the exact asymmetries the model uses to keep you recognizable, so the output drifts toward a generic face. Group shots force the model to pick a person, and it often picks wrong. Side profiles and downward angles hide half the face, and the model invents the rest.

A phone selfie in daylight against a plain wall is close to ideal. You do not need a studio; you need a face the model can actually see.

2. Give a topic the lyrics can use

A rap needs something specific to be about. "Summer" gives the model almost nothing, so the bars come out generic. "The summer I moved to a city where nobody knew my name" gives it a scene, a feeling and a detail, and the lyrics get sharper for it.

If there is a name, a place or a punchline you want in the rap, put it in the topic. The model writes around what you give it, and a specific input is the difference between a clip that feels personal and one that feels like filler.

3. Choose a length

Length is a trade-off between completeness and risk. A 5-second clip is a hook: fast to render, cheap, and short enough that the model never has time to drift. A 15-second clip is a full short verse with a hook, but it gives the model more frames to keep your face consistent, so it is less forgiving of a weak photo.

One credit equals one second of video, so a 5-second clip costs 5 credits and a 15-second clip costs 15. If you are testing a photo, render a 5-second version first. If the face holds, spend the credits on the longer one.

4. Confirm consent and generate

Before the render starts you confirm that the photo is yours or that you have permission from the person in it, and that they are an adult and not a public figure. This is not a formality; it is the line between a creative tool and a tool for making non-consensual content.

Once you confirm, the render runs in the background and usually finishes in two to eight minutes. You can close the page. The finished video waits for you in My videos, and if the render fails, the credits go straight back to your balance.

5. Download and share

When it finishes, download the MP4. The file has no watermark and is already vertical 9:16, so it drops straight into Reels, Shorts or TikTok without cropping.

The clip is labelled as AI-generated on the site, and the label matters when you post it: say it is AI when you share it. That is both the honest thing to do and the thing that keeps the format usable for everyone.

Why some results come out fake

Almost every bad result traces back to one of five causes. The photo hid the face (sunglasses, filters, angles). The photo did not match the expression the clip needed, so the model invented one. The lighting of the photo fought the lighting of the scene, which reads as wrong even when the face is clean. The clip was long enough for the face to drift. Or the topic was too vague for the lyrics to carry.

Fixing these is not about finding a better setting. It is about giving the model a clear face, a specific topic, a short enough clip, and lighting that belongs to the scene. Do those four things and the render has everything it needs.

Why the first two seconds matter most

In a feed, the viewer decides whether to stay within the first two seconds. That means the opening frame has to do the work: face visible, expression alive, motion already underway. A version that builds slowly and saves its best frame for second four loses the viewer before the reveal.

When you are choosing which take to post, judge it by the first two seconds, not the last. If the face does not read as a real person immediately, no amount of good material later will recover the watch time.

How to write a topic that produces good bars

A topic works when it contains a scene, a feeling and at least one concrete detail. "My first apartment" is thin. "The first apartment where the radiator knocked all night and I still could not sleep without it" is a topic, because it gives the lyric a sound, a place and a contradiction to play with.

The most common mistake is writing a category instead of a moment. Categories ("summer", "money", "friends") produce lines that could belong to anyone. Moments produce lines that could only belong to you, and that is the whole point of putting your own face in the video.

If you get stuck, write the topic as a sentence you would actually say out loud to a friend. That sentence usually contains the hook already; the model just needs to hear it.

How credits and lengths actually line up

Because one credit equals one second, the length you choose is the amount you spend. A 5-second test costs 5 credits, a 10-second version costs 10, and the full 15-second version costs 15. Packs are bought once and the credits never expire, so testing a photo on a short render costs almost nothing relative to the final video.

The practical rhythm is: test short, then commit. Render a 5-second version of a new photo before you spend on the full length. If the face holds for five seconds, it will almost always hold for fifteen. If it drifts in five, a longer render will only drift more, and you will have spent three times the credits to learn the same thing.

If a render fails, the credits go back to your balance automatically, so a failed attempt does not cost you anything except time.

Common mistakes to avoid

Uploading a photo you have already filtered for social media. Copying a clip you saw and expecting the same framing. Rushing the topic. Rendering the longest length first on an untested photo. Re-rendering the same broken setup and expecting a different result instead of changing the photo. Each of these wastes credits, and each is easy to avoid once you know which variable actually controls the outcome.

How to judge your own result

Watch the finished clip once with the sound off, then once with the sound on. With the sound off, ask only one question: does this read as a real person doing something, or as a face on a moving image? If the answer is the second, the problem is almost always the photo or the length, not the lyrics. With the sound on, check whether the bars say anything only you could say. If both pass, the clip is worth posting; if only the first passes, the topic needs another pass before the next render.

The order matters. Fixing the face is cheaper than fixing the lyrics, because the face is a photo decision and the lyrics are a prompt decision, and a re-shoot of one photo costs nothing while a re-write of a topic costs a full render cycle. Judge the face first, then the words.

What to do next

Pick a clear, front-facing photo, write down the one thing you want the rap to be about, and render a 5-second version to test it. If the face holds, make the longer one. If you want the background on why this format spread in the first place, read the trend guide first.