Prompt in your delivery format from the first line, block for a phone screen where the face lives in the top third, and route close-ups to Veo and multi-shot dialogue to Kling — all inside the invideo agent, which enforces the vertical frame across the storyboard and DOP agents.
Vertical AI micro-drama shots come out right when you prompt in 9:16 from the first line, compose for a phone screen — face in the top third, negative space at the bottom for captions — and route each shot to the model that handles it best: Veo for close-ups, Kling for multi-shot dialogue, Seedance 2.0 when a character or spatial reference must carry across clips. Inside the invideo agent, the vertical frame is set once and enforced across your storyboard and DOP agents. This page covers shot generation specifically; for the full production workflow, start with our complete guide to AI micro-drama.
Why vertical shot generation needs different prompting
Video models are trained predominantly on horizontal footage — cinema, television, landscape web video — so their compositional defaults are landscape defaults: subject centered at mid-height, headroom balanced for a wide frame, action spread left to right. Set a vertical aspect ratio without changing anything else and you get a landscape composition squeezed into a tall frame: the face lands mid-screen, dead space stacks above the head, and the action pushes out both sides of the crop.
Most AI video generation tools accept 9:16 as an output parameter; that only changes the shape of the canvas, not how the model composes inside it. Vertical has to be a directing decision written into every prompt — where the face sits, what fills the bottom of the frame, how depth is layered top to bottom. If you're using an AI short film generator for a phone-first series, that decision compounds across dozens of shots per episode, so it belongs in the project setup, not in per-shot improvisation. invideo is an agentic video creation tool with all the current video models available, and when you set the vertical format at project level, the storyboard agent and DOP agent inherit it for every shot they plan and generate — the frame stays enforced without you restating it.
Frame for the phone screen: composition rules
Compose every vertical micro-drama shot for the device it plays on, not for an abstract tall rectangle. The rules that hold across shot types:
- Face in the top third, negative space at the bottom for captions. Put the eyeline near the top-third line; the bottom third stays visually quiet so burned-in dialogue text and platform captions never cover a performance.
- Respect the UI safe zones. Keep critical action away from the bottom caption band, the right-edge engagement buttons, and the top status bar — a reaction staged behind a like button is a wasted generation.
- One subject per frame. A 9:16 frame doesn't have the width for a conventional two-shot; stage dialogue as alternating singles and let the cut carry the exchange.
- Layer depth vertically. Where a landscape frame spreads foreground and background left to right, a vertical frame stacks them: foreground element low in frame, subject above it, environment behind.
For high-stakes scenes, lock the composition before spending video credits: lay out the key frames as a single vertical composite image in GPT-Image-2, then have the invideo agent attach that storyboard to Seedance 2.0 and animate from it. As invideo's production guidance puts it: "For high-stakes scenes - Storyboard first. Lay out the key frames as one vertical image. The Agent attaches it to Seedance and animates from it. Lower credits spent, higher control." You approve the vertical framing on stills, where corrections are cheap, and the animation inherits it.
Prompting each shot in your delivery format
Write the frame into the prompt itself rather than relying on the aspect setting alone. Specify the orientation and then describe what occupies the top, middle, and bottom of the frame: "vertical 9:16 frame, close-up, her face in the top third, rain-streaked window filling the lower two thirds" produces a phone-native composition; "close-up of her face" produces a landscape close-up that happens to be tall. The framing language does the work — subject position, headroom, what fills the negative space — because that's what overrides the model's horizontal training bias shot by shot.
Two generation habits keep vertical dialogue scenes usable in the edit. First, don't cram a full exchange into a single 15-second generation — it breaks pacing and reduces your editor's control; split the scene into single-beat shots and let the cut set rhythm. Second, generate 3–4 options per shot so the edit has coverage to choose from rather than one take to accept.
Set the format once in the invideo AI video generator and every shot prompt runs against it — the invideo agent carries the vertical spec, the composition rules, and your locked references into each generation, so the per-shot prompt only has to describe the shot. As one production team described the shift after working this way: "After a while, it stops feeling like prompting and it just feels like directing."
Which model handles vertical best per shot type
Route vertical shots by type rather than running everything through one model — each current model has a distinct strength in the 9:16 frame:
- Veo — close-ups. Strongest on facial performance in the tight vertical single, which is the dominant shot in phone-first micro-drama: the face holds the top third and the expression carries the beat.
- Kling — multi-shot dialogue. Generates multi-shot sequences natively, which suits the alternating-singles structure vertical dialogue demands; you get connected coverage of an exchange instead of stitching unrelated generations.
- Seedance 2.0 reference-to-video — continuity shots. Attach a previous shot or a vertical storyboard frame as a spatial layout reference and the generation keeps the same space, lighting, and character position — the tool for any shot that must match what came before. For carrying a locked character across clips, see how we handle character consistency.
The reference-to-video route also gives you a sequencing method for whole scenes: start every sequence with a wide establishing shot, then have each subsequent shot reference the last through Seedance 2.0 — "The Agent holds the spatial layout. Every shot after references the last - same space, same lighting, same character position." For insert shots and passing montage beats, skip location reference generation entirely and prompt the model directly from the script description; the continuity machinery isn't worth the credits on a two-second cutaway.
All of these models run inside invideo, so routing is a per-shot decision, not a platform decision — the invideo agent assigns each shot to the right model from your breakdown. The full decision logic is covered in model routing by shot. The approach holds up at series scale: one documented 3-person team shipped a 10-episode vertical micro-drama series in 3 days at $1,000 per episode using exactly this per-shot routing inside the invideo agent.
Reframing horizontal generations when a model insists
When a generation comes back composed horizontally despite a vertical spec, you have three fixes, in ascending order of cost:
- Crop a vertical slice. Punch into the horizontal frame and extract the 9:16 region around your subject. This works when the subject holds one zone of the frame for the whole clip; it fails when the action traverses the width. You're discarding resolution, so budget an upscale pass on the cropped result.
- Extend the frame vertically. Pull the best frame from the horizontal clip, extend the canvas top and bottom in GPT-Image-2 so the composition becomes natively vertical, then animate the extended frame through Seedance 2.0. You keep the performance and lighting you liked and gain the vertical real estate the model refused to generate.
- Regenerate with a vertical reference. Attach a vertical storyboard frame or a prior approved vertical shot through Seedance 2.0 reference-to-video and regenerate. The reference forces vertical composition from the start, and it's usually cheaper than repeatedly cropping and repairing outputs that were composed wrong.
If a specific shot type keeps coming back horizontal, move that shot to the storyboard-first path from the composition section above — approving a vertical still before animation removes the failure mode entirely.
FAQ
How do you generate vertical AI video?
Set 9:16 at the project level, then describe vertical composition in every prompt — subject position, what fills the top and bottom of the frame — rather than relying on the aspect parameter alone. Compose with the face in the top third and negative space at the bottom for captions. Inside the invideo agent, the vertical format is set once and enforced across the storyboard and DOP agents for every shot.
Which AI video model is best for 9:16?
Route by shot type: Veo for tight vertical close-ups and facial performance, Kling for multi-shot dialogue sequences, and Seedance 2.0 reference-to-video for any shot that must match a previous shot's space, lighting, or character position. All of these models run inside invideo, so the choice is made per shot, not per platform.
How do you fix a horizontal AI generation?
Crop a 9:16 slice when the subject holds one zone of the frame, extend the frame vertically in an image model and re-animate it through Seedance 2.0 when you want to keep the take, or regenerate with a vertical reference attached so the model composes vertically from the start. If a shot type repeatedly comes back horizontal, storyboard it as a vertical still first and animate from the approved frame.