A good AI rap video looks simple: one person, one beat, sixteen seconds of confident delivery. Getting there is less simple. The model has to keep the same person on screen from the first frame to the last, move them in a way that reads as one continuous take, and line up their mouth with words that did not exist a minute ago.
Press enter or click to view image in full size
This tutorial walks through how to do that with MiniMax H3 (Hailuo 3.0), a video model that generates picture and sound in the same pass. It covers:
What H3 can and cannot do for a rap video
How to prepare the photos you upload
How to structure the prompt
How to keep the character consistent
How to keep the motion path continuous
How to get usable lip sync
A complete example prompt
A troubleshooting checklist

Everything here applies whether you call H3 through an API or through a tool built on top of it. The examples assume one rapper, but the same rules scale to duets and group shots.
- What H3 Gives You (and the Limits to Plan Around)
H3 accepts text plus reference material and returns a short clip with synchronized audio. For rap videos the relevant facts are:
Item What to know Clip length 5–15 seconds per generation, 24 fps. Longer pieces are built from several clips. Prompt length Up to 7,000 characters. Longer prompts are rejected, not truncated. References Up to 12 in total: 9 images, 3 videos, 3 audio files. Audio Generated together with the picture, so vocals, beat and mouth movement come from one pass. Negative prompt There is no separate field. Exclusions go inside the main prompt.
Two consequences follow. First, plan your video as a set of 10–15 second beats rather than one long take. Second, because sound and picture are generated together, anything you write about the music can affect the body, and anything you write about the body can affect the music. You will see this again in the lip-sync section.
- Preparing Your Photos
The photo is the single biggest factor in whether the result looks like the person you intended. H3 can only preserve what it can see.
Use photos you have the right to use
Only upload photos of yourself, or of people who have agreed to be in the video. Do not use photos of celebrities, public figures or anyone who has not said yes. This is not just etiquette: most platforms that host or process AI video, and the payment providers behind them, will remove content or close accounts over it.
What makes a good reference photo
Requirement Why it matters One person per photo If two people are in frame, the model has to guess which one you mean. Crop to the person or use a solo shot. Face clearly visible, roughly front-facing A three-quarter angle is fine. Profiles, heavy shadows and hands over the face lose detail the model then invents. Face large enough in the frame Head and shoulders, or waist up. A full-body shot from across the room leaves the face as a few hundred pixels. Even, soft light Harsh side light bakes shadows into the identity; the model may keep them even when the scene is lit differently. Sharp, not filtered Avoid beauty filters, heavy compression, motion blur and stickers. Filters change face shape, and the model will faithfully copy the filtered version. No sunglasses, masks or hats that hide the hairline Unless you want them in the video. Whatever is in the photo tends to stay. Clothing you want to see on screen The outfit in the photo is the outfit the model is most likely to keep. Reasonable resolution A recent phone photo is plenty. Very small images (for example, a cropped avatar) give the model little to work with.
One photo or several?
One clear photo is usually enough for a single performer. Add a second photo when:
the scene will show an angle your first photo does not (a side view for a profile shot, a full-body photo for wide framing), or
the outfit matters and the first photo is cropped too tight to show it.
When you upload more than one photo of the same person, say so in the prompt and say which one wins when they disagree (for example, “Image 1 controls the face, Image 2 controls the outfit”). Otherwise the model may blend them.
Press enter or click to view image in full size
Choosing a reference video
If you want the rapper to follow a specific performance — a walk toward camera, a hand gesture on the downbeat, a turn on the hook — a short reference video is far more reliable than describing it in words. Pick a clip that:
has the same number of people you want in the result,
has a steady camera or a camera move you actually want,
has no hard cuts inside the part you use, and
matches the length you plan to generate.
You will tell H3 to take the motion and camera from this video and nothing else.
- Structuring the Prompt
H3 responds best to a prompt organized into labeled sections, each with one job. The six-field layout below is a reliable default for rap videos:
Field What goes in it subject_definitions Give every input and every person a fixed name (Image 1, Video 1, Subject 1) and say what each one is for. Use these names everywhere after this. summary One paragraph: what the clip is and how the inputs relate to it. The model weighs this most, so put your most important rule here once. retention_analysis For each reference, how much to keep and which attributes. detailed_description Visual style first, then the clip in time order: action, camera, and sound tied to moments. Character rules, props, background and constraints go here as short sub-sections. overall_soundscape Ambient and physical sound across the whole clip: room tone, crowd, footsteps. non_diegetic_music Music the audience hears but that has no source in the scene, such as the beat — or a note that there is none.
If you prefer plain prose, keep the same order: references → what to retain → scene → timeline → camera → audio → constraints.
Rules that make a prompt work
Assign roles at the top. Start with “Use Image 1 as…” and “Use Video 1 as…”. Explicit roles are followed more reliably than files that are merely mentioned.
Use retention levels, and pick them carefully. In retention_analysis, label each reference:
fully_preserved — keep it as it is (for example, a background plate).
partially_preserved — keep only the listed aspects (for example, a reference video's motion and camera, but not the person in it).
attribute_transfer — move listed attributes onto a new subject (for example, the face, hair and outfit from your photo onto the performer).
weak_reference — mood or style only.
If a reference contains something you want to keep and something you do not, never mark it fully_preserved. A reference video used for choreography should be partially_preserved, with a sentence saying none of the original performer's appearance carries over.
Write timestamps. Use [0–3s] for a span and [5.4s] for a moment. Sound can be pinned to a moment too: [6.2s] The bass drops and he points at the lens.
Use film language for the camera. “Slow dolly in”, “locked-off medium shot”, “handheld track left”, “whip pan”. Describe the camera in its own sentence, separate from what the person does.
Write exclusions as sentences, and only the ones you need. “No extra people, no on-screen text, no logos, no morphing.” Every word in an exclusion list reads as a ban, so a list that mentions subtitles will suppress subtitles even if another line asks for them.
Download the Medium app
Keep it coherent. A long prompt is fine; a contradictory one is not. “Keep the camera exactly as the reference” followed later by “feel free to reframe” gives the model two jobs and it will do neither well.
- Keeping the Character Consistent
Identity drift — the face slowly changing, the jacket turning a different color, a tattoo appearing and vanishing — is the most common way an AI rap video falls apart. Three habits prevent most of it.
Name the features, don’t just ask for “consistency”
“Keep the character consistent” gives the model nothing to check against. List what must hold:
Subject 1 keeps Image 1’s face shape, eye shape, nose, skin tone, facial hair, short twisted hairstyle, gold hoop in the left ear, black oversized hoodie with white drawstrings, and slim build for the entire clip.
Specific nouns give the model anchors. Generic adjectives (“cool outfit”, “stylish hair”) give it room to improvise.
Give each person one name and never change it
If you call the rapper “Subject 1” in one section, “the man” in another and “the performer” in a third, the model can read them as three people. Define the name once in subject_definitions and reuse it exactly.
Account for everyone in the frame, including people facing away
If your reference video has two people and one of them is turned away from the camera, say explicitly that the person facing away is still a character slot and still needs the new hair, back of the head, shoulders and clothing. If you don’t, the model tends to change only the faces it can see, and you end up with one new person and one leftover from the reference.
Frame for the face
The smaller the face on screen, the less the model can hold on to. Medium shots and closer keep identity best. If you want a wide shot, keep it short and cut back to a medium shot for the lyric-heavy part.
Across several clips
When a video is built from multiple generations:
reuse the same photos and the same identity paragraph, word for word, in every clip;
keep wardrobe and lighting descriptions identical unless the change is the point;
carry the end state of one clip into the opening of the next (see the next section).
- Keeping the Motion Path Continuous
A clip feels like one take when every movement has a clear start, a path and an end, and when the end of one beat is the start of the next. Most “glitchy” motion — limbs snapping into place, a sudden jump in position — comes from prompts that describe poses instead of movement.
Describe trajectories, not end poses
Weak:
He raises his hand.
Strong:
[3–5s] Starting with both hands at his chest, he lifts his right hand slowly to shoulder height, palm out, and holds it there through the end of the line.
Start state → movement → end state. The model fills the frames between two well-defined positions far better than it invents both ends.
Chain the beats
Write each beat so it begins where the last one ended:
[0–4s] Subject 1 walks toward the camera from the doorway and stops at the center mark, weight on the left foot. [4–9s] From the center mark, still facing the lens, he starts the verse, shoulders relaxed, small nods on the snare. [9–12s] On the hook he steps half a step to his right and lifts his right arm, then lowers it back to his side by [12s].
No beat starts somewhere the previous beat did not leave him.
Separate camera motion from body motion
If the camera and the body are described in the same sentence, the model may move one when you meant the other. Give the camera its own line:
Camera: locked-off medium shot at chest height for the whole clip; a slow dolly in begins at [9s] and ends at [12s].
Let a reference video carry the hard parts
For anything with real choreography, a reference video beats any description. Mark it partially_preserved, list exactly what to keep — "motion path, timing, body position, gestures, framing and camera movement" — and list what not to keep — "the original performer's face, hair, body and clothing". Then tell the model the motion must match the reference frame by frame, not just "be similar".
Don’t cut inside a clip unless you mean it
If you want one continuous shot, say so: “One continuous shot, no cuts.” H3 is happy to insert cuts on a beat drop, and a cut resets the motion path.
Joining clips
H3 tops out at 15 seconds, so longer videos are stitched. To make the joins invisible:
End each clip on a calm, held pose rather than mid-gesture.
Describe that exact end pose (position, facing, hand placement, camera framing) at the start of the next prompt.
If your tool supports it, use the last frame of the previous clip as an image reference for the next one.
Keep the camera framing the same across the join; change framing on a musical accent instead, where a cut feels intentional.
- Getting Lip Sync That Reads as Rapping Put the lyrics in quotation marks
H3 treats quoted text as words to be spoken or sung and syncs the mouth to them:
[4–9s] Subject 1 raps: “Woke up early, city still asleep, every step I take is a promise I keep.”
Writing “he raps about the city” produces a mouth that moves vaguely; quoting the line produces a mouth that forms the words.
Fit the words to the time
Rap is dense, but not infinitely so. As a rough guide, budget two to three words per second for clean delivery. If a line is too long for its time slot, the model will rush it, slur it or cut it off. Shorter lines with clear rhythm sync better than long ones.
Keep the body out of the lip sync
Because H3 generates the vocal and the picture together, it often lets the music drive the whole body: heads bob, shoulders bounce, hands start dancing, and the motion you carefully specified is overwritten. Fix this by saying which parts may change for the vocal:
Only the mouth and jaw move to form the lyrics. Head position, upper body, hands and the motion path follow the reference video exactly. Do not dance or bounce to the beat unless the reference does.
Put that rule at the start of the motion section and repeat it once in the summary.
Avoid instructions that contradict each other, such as “the music must not change his mouth” — that asks for new lyrics without mouth movement. And avoid phrasing like “his body rides the rhythm” unless you actually want the body moving with the beat.
Describe the music separately from the vocal
Put the beat in non_diegetic_music (tempo, genre, instruments, energy) and the vocal in the timeline. For example: "A 90 BPM boom-bap beat with a dusty drum loop and a muted piano sample; no other vocals." The vocal itself lives in the quoted lines.
- A Complete Example
Here is a full prompt for a 12-second, single-performer clip using one photo and one reference video. Adapt the names, lyrics and setting.
subject_definitions:
Image 1 is a photo of Subject 1, the performer, who has agreed to appear in this video.
Video 1 is a reference clip of one person performing to camera on a rooftop at dusk.
Subject 1 is the only person in the output.
summary:
Create a 12-second rap performance in which Subject 1, with the appearance from Image 1, performs on the rooftop from Video 1, following Video 1's motion and camera frame by frame. Only Subject 1's mouth and jaw change to form the new lyrics; head, body, hands and the motion path follow Video 1 exactly. One continuous shot, no cuts.
retention_analysis:
Image 1: attribute_transfer. Preserve face shape, eye shape, nose, skin tone, short twisted hair, gold hoop in the left ear, black oversized hoodie with white drawstrings, slim build.
Video 1: partially_preserved. Preserve rooftop background, dusk lighting, motion path, timing, body position, gestures, framing and camera movement. Do not preserve the original performer's face, hair, body or clothing.
detailed_description:
Style: cinematic music-video look, warm dusk backlight, soft city bokeh, natural skin texture.
[0–4s] Subject 1 walks from the rooftop edge toward the camera and stops at the center, weight on the left foot, hands loose at his sides. Camera: locked-off medium shot at chest height.
[4–8s] From the center, facing the lens, Subject 1 raps: "Woke up early, city still asleep, every step I take is a promise I keep." Small nods only where Video 1 nods.
[8–12s] On the hook he lifts his right hand from his side to shoulder height, palm out, as in Video 1, and raps: "Hold it down, never let it go." He lowers the hand back to his side by [12s]. Camera: slow dolly in from [8s] to [12s], ending on a medium close-up.
Identity: Subject 1 keeps every attribute listed for Image 1 in every frame, including when turned slightly away from the camera.
Motion: only the mouth and jaw move for the lyrics. Do not dance or bounce to the beat beyond what Video 1 shows.
Constraints: No extra people. No on-screen text or logos. No cuts. No morphing of the face or clothing.
overall_soundscape:
Light rooftop wind, distant city traffic, faint footsteps during the walk-in.
non_diegetic_music:
A 90 BPM boom-bap beat with a dusty drum loop and a muted piano sample. No other vocals.
Notice what the example does: every input has one name and one job; the most important rules (appearance from Image 1, motion from Video 1, only the mouth changes) appear in the summary and again where they apply; each beat starts where the previous one ended; the camera has its own line; and the exclusions are short and specific.
Writing a prompt like this from scratch every time is the slow part. One shortcut is to save a working prompt as a template and change only the lyrics and photo between runs — that is the approach behind template-based tools such as ai rap video, where the motion, camera and constraints are fixed per template and you supply the photo and the topic. Whether you use a tool or your own saved prompts, the principle is the same: lock down everything that already works, and vary one thing at a time.
- Troubleshooting Checklist
Symptom Likely cause Fix The face looks like someone else Low-quality or filtered photo; face too small on screen Use a sharp, unfiltered, front-facing photo; frame medium or closer. Part of the reference performer is still visible Reference video marked fully_preserved Use partially_preserved and state that none of the original appearance carries over. A person facing away was left unchanged Prompt only described visible faces Say that people facing away or partly hidden are still character slots and need new hair, back of head, shoulders and clothing. Outfit changes mid-clip Outfit not named specifically List garments, colors and accessories in the identity paragraph. Body starts dancing to the beat Music driving the body State that only the mouth and jaw change; repeat in the summary. Limbs snap or teleport Poses described instead of movements; beats don't chain Write start → path → end for each move; begin each beat where the last ended. Unwanted cut on the beat drop No continuity instruction Add "One continuous shot, no cuts." Mouth doesn't match the words Lyrics paraphrased, or too many words for the time slot Quote the lyrics; aim for two to three words per second. An element you asked for never appears The same word appears in the exclusion list Remove it from the exclusions; describe it positively with position and timing. Request rejected outright Prompt over 7,000 characters Cut repetition first; keep each rule in one place plus one reminder in the summary.
Final Notes
Change one thing per generation. If you change the lyrics, the photo and the camera move at once, you won’t know which change fixed or broke the clip.
Save every prompt that works. Your best prompts become your templates.
Keep consent at the center. Use your own photos or photos of people who agreed, and don’t present generated clips as real footage of anyone.
With clean photos, a structured prompt, explicit identity anchors and motion written as connected trajectories, H3 can produce rap clips that hold up on a phone screen — which is where most of them will be watched.

Top comments (0)