Cinematic Mode: AI Film Scenes for Music Ads That Do Not Look Like AI Slop

Cinematic mode writes a short fictional film scene from your song: two characters, one location, three shots with hard cuts, dialogue and a film look. The story comes from the audio analysis, the shot grammar is fixed so the camera work stays consistent, and each shot can be regenerated on its own.

By the song.so team

Why a scene instead of a slideshow

Most AI music ads look alike: a still image with a slow zoom. Viewers have seen thousands of them, and the sameness has a name now, AI slop. A scene with two characters, real coverage and cuts reads as a film, and that is what stops a thumb in Reels and Stories. Cinematic mode is built around that idea: one location, two people, three shots, and a look that avoids the clean digital finish.

It is not trying to produce a feature film. A scene lasts about six seconds, which is the length of the part of a song you can fit in a short ad. In that time it has to establish a place, introduce two people and land one small story beat that connects to the music. That constraint is what makes it work. Short, specific and visual beats long and vague.

How the scene is written

The audio of your snippet is analyzed for genre, mood, topics and vibe. A scene director then writes a scene bible: the location, two characters with age and look, a music-related action and one story beat per shot. You can add a reference image to fix the environment and possibly one object or character, and you can give a story direction in your own words. The reference is read first and turned into a compact written description, so the story call that follows stays fast and every image shows real progress.

The result is a set of still-frame prompts and a motion prompt for each shot. The model that writes the story is chosen from several, and if one fails, the next is tried, and the story step is retried several times before the job gives up. That is a quiet reliability feature: you rarely see a failure caused by one model being busy.

By default a scene has two people. If your story direction clearly asks for one person only, for example someone alone, sleeping or on their own, the scene becomes a solo scene with its own shot pattern: a wide shot, a close-up and a medium shot from a different angle. The system also recognizes when you describe a crowd or onlookers and keeps them in the background as passive extras, rather than collapsing a group into a single second character.

A fixed shot grammar

The model writes the story but not the camera rules. The shot grammar is deterministic, so coverage does not depend on the mood of the model. For two-person scenes there are two patterns: a wide two-shot followed by an over-the-shoulder shot on character A and then on character B, or an over-the-shoulder shot on A, a wide shot, and an over-the-shoulder shot on B. Character A is always on the left and character B on the right of the 180 degree axis, and the last frame is always the reverse shot on B.

PatternShot 1Shot 2Shot 3
Two-person, wide firstWide two-shotOver the shoulder on AOver the shoulder on B
Two-person, close firstOver the shoulder on AWideOver the shoulder on B
Solo, wide firstWide establishing shotTight close-upMedium shot from another angle
Solo, close firstTight close-upWide establishing shotMedium shot from another angle

The reason for a fixed grammar is craft. Wide shots establish the place, close-ups carry emotion, and the 180 degree rule keeps the viewer oriented. Without those rules a model tends to produce three similar medium shots that feel like three photos instead of a scene. The wide shot is written to show both characters head to toe with the floor visible, the close-ups are tight with a shallow depth of field, and the over-the-shoulder shots show the other person only as a soft foreground shoulder.

From frames to a finished clip

Each shot starts as a still frame, and the three frames are captured first with the film look already built in. Every frame gets up to three attempts: the first attempt and up to two corrections, at the same charge, and the attempt with the fewest problems wins. You can regenerate any frame, compare versions and pick the best one, and you can edit a single frame with a text instruction and accept or reject the result before it replaces the original.

When you edit one frame, a free consistency check compares every other frame with the edited one and reports which no longer match, looking at people, wardrobe, accessories, props and location. A different camera angle or framing is expected and never counts as a mismatch. The check is designed to fail open, so an error in the check never blocks you.

Every frame is then animated into its own six second clip, because a single video call ignores hard cut instructions and produces one continuous take. The clips are joined with real cuts into a six second scene. Each shot gets an equal slice, about two seconds, so the cuts land at around two and four seconds. For each clip you choose which part to use, the start, the middle or the end of the animation, and you can reorder the shots and edit the action prompt of each one before you finish. Nothing is merged until you have made those choices.

StageWhat happensWhat you control
StoryScene bible: location, two characters, a beat per shotStory direction, reference image
FramesThree stills with the film look baked inRegenerate, edit, compare versions
ConsistencyFree check of the other frames after you edit oneFix any frame that no longer matches
AnimationOne clip per frameAction prompt per shot
MergeHard cuts into a six second scenePart per shot (start, mid, end), shot order
FinishAudio laid under the scene, text addedHooks, lyrics, overlay, film look

It keeps working when you close the tab

Frame capture and the optional grade pass run as a background job, independent of the browser. The page only watches the job, and a scheduled worker moves it forward every minute, so it finishes even if the tab is closed. Each external task is checked rather than waited on, with a hard cap of around twelve minutes, and a frame that cannot be completed is reported instead of hanging forever. This matters in practice, because generating three frames and animating three clips can take a few minutes, and nobody wants to stare at a progress bar.

The look

The prompts ask for a photochemical image: anamorphic lens character, shallow depth of field, practical light, high contrast with deep blacks, visible film grain, halation and natural motion blur. They also forbid digital sharpening, HDR and face smoothing, so skin keeps its pores and imperfections. The frames are captured with the look of a 35 mm cinema camera already built in. An optional grade pass can re-apply the same look to a finished frame, and a film emulation can be added on top of the final video, as explained in film emulation and color grade.

Why does this matter for ads? Because the clean, evenly lit, over-sharpened look is exactly what people now associate with generated content. Grain, contrast and imperfect skin are cues that read as a camera, and viewers give a camera more attention than a render.

Writing a good story direction

You can leave the direction empty and let the audio analysis decide, or give a short brief. A few habits produce better scenes. Name the place and the situation, not the plot: a laundromat at night, two strangers on a night bus, a barber shop on a quiet afternoon. Keep it to one idea. Say who the two people are in a few words. Mention an object if you want one in every shot. And leave the music out of it, because the scene director already links the action to the sound.

Weak directionBetter direction
An exciting scene with lots of actionTwo friends in a parked car at night, one shows the other a song
A story about love and lossA woman alone in a kitchen at 2 a.m. listening on one earbud
Make it cinematicA barber and a customer in an empty shop, the radio is on

If you want one person only, say so plainly, for example a person alone. If you want a crowd, describe it as watching in the background, so it is treated as scenery.

Step by step: making your first scene

The first scene is the best way to learn the mode. This is the sequence.

StepWhat you doWhat to look for
1Choose a snippet with a clear moodA section that already feels like a moment in a film
2Choose Cinematic as the modeThe flow switches to the scene steps
3Add a reference image and a short story direction, or leave both emptyA reference fixes the place, a direction fixes the idea
4Let the three frames generateTwo characters, one place, three different framings
5Check the frames, regenerate or edit any that are offSame people, same clothes, same location in every frame
6Animate the clips and choose a part and an order for each shotPick the start, middle or end of each clip that looks best
7Finish, then add a hook, lyrics and the cover overlayA hook that names the situation

Expect to regenerate a frame or two on the first attempt. That is normal, and the process is designed for it: you pick the best of the versions, and the consistency check tells you if one frame has drifted from the others.

Story directions by genre

The scene director reads the genre and mood from the audio, but a short direction steers it toward a setting that fits. These are starting ideas, not recipes.

Genre or moodSetting ideaWhy it fits
Late night electronicTwo people in an empty parking structure, a phone playing the trackUrban, nocturnal, high contrast
Warm indie or folkTwo friends on a porch at golden hour, a speaker between themSoft light, natural color
Hip hopA barber shop or a stairwell, one person shows the other a songIntimate, local, real
PopA kitchen or a car, someone discovers the song and reactsRelatable, bright, quick payoff
Melancholic balladA single person on a night bus, one earbud inSolo shot pattern, quiet mood

Using scenes in ads

A scene is one creative with a story, so use it as a contrast to your performance clips and hooks. Add a text hook, lyric captions and the cover overlay, then export variations. A scene with a hook that names the situation often outperforms the same scene with no text, because it tells the viewer where to look. For the strategy behind mixing formats, see native content beats music videos and bulk ad creative variations.

Make several scenes rather than one. Meta rates the diversity of the ads in an ad set, and three different scenes with different places and characters are far more diverse than three versions of the same scene. A realistic goal for a campaign is two or three scenes next to your other formats.

FAQ

How many shots does a cinematic scene have?

Three, with hard cuts at about two and four seconds, joined from three separate clips into a six second scene.

Where does the story come from?

From the audio analysis of your snippet: genre, mood, topics and vibe. A scene director writes the location, two characters and one beat per shot. You can also give a story direction and a reference image.

Can I control the scene?

Yes. You can add a reference image, regenerate single frames, choose between frame versions, edit the action prompt per shot, set the part of each shot and change the shot order.

Why are shots generated separately?

A single video generation call ignores hard cut instructions and produces one continuous take. Separate clips joined together give real cuts.

Can I add a film look?

Yes. The frames already carry a film look, and a film emulation with grain, halation, vignette and more can be applied to the final video without affecting the captions.

Can I make a scene with only one person?

Yes. If your story direction asks for one person only, the scene uses a solo shot pattern with a wide, a close-up and a medium shot.

Do I have to keep the page open?

No. Frame capture and grading run as a background job that keeps going when the tab is closed.