Two people talking should be the easiest scene in the world to generate. It is not. Ask a model for a conversation and you usually get one locked-off wide shot, two figures facing the camera like news anchors, mouths moving out of sync with words nobody can quite hear, and a slow drift where one actor’s face quietly becomes somebody else halfway through the sentence. I’ve burned more credits on simple two-person dialogue than on any action scene I’ve ever attempted. The lesson finally stuck: AI dialogue doesn’t fail because the models can’t render talking people. It fails because most creators generate a conversation the way they’d film an interview, when they should be generating it the way editors cut one.
Real film dialogue is rarely one shot. It’s coverage — a set of angles filmed separately and assembled so the audience feels they’re watching two people react to each other in real time. The good news? Coverage is actually friendlier to AI generation than a single master shot, because each clip only has to do one small job. One person speaks, one person listens, nobody has to share a frame and stay consistent for twenty seconds. This guide breaks down the coverage kit I use for generated conversations, how to prompt each piece, how to keep eyelines from making your actors stare into the void, and how to cut it together in Premiere Pro so it feels like a scene instead of a slideshow.
Why Single Wide Shots Kill Dialogue
A single wide of two actors carries four problems at once, and any one of them can sink the scene. First: lip sync at wide framing is unforgiving in the wrong way — the mouths are small enough that sync errors blur, but the body language is fully visible, and generated bodies rarely act. They sway. They gesture half a beat late. They hold eye contact with the lens instead of each other. The audience can’t name what’s wrong, so they just feel it.
Second, the wide gives you nowhere to hide in the edit. If the performance in seconds four through seven is weak — and in a long generated clip, some stretch always is — you can’t cut away, because there’s nothing to cut away to. You either keep the bad stretch or scrap the whole generation. Coverage turns that problem inside out: every weak moment in one angle can be covered by a reaction from the other angle. Which is exactly how human editors have rescued human actors for decades.
Third, two characters in one frame doubles your consistency risk. Each face gets fewer effective pixels, the model splits its attention, and wardrobe details start migrating between the two figures. Put each actor in their own shot and each generation has one job: keep this one person looking like themselves. That’s a much easier job. And the consistency techniques in the guide to character consistency in AI video work noticeably better when a frame holds a single face.
And honestly? The wide is boring to look at for more than a few seconds. Dialogue scenes live on changing viewpoints. The audience wants to be near the person speaking, then near the person absorbing it. A static wide denies them both. And on short-form, denied viewers leave.
The Coverage Kit: Wide, Two OTS, Two Reactions
For a two-person scene, I storyboard five shots. Not because five is magic — but because this kit covers nearly every conversation you’ll ever cut:
- The wide (or master). Both characters, full environment. Use it for the first two or three seconds to place the scene, and again at the end if the scene needs air. Keep it short; it is a map, not the territory.
- Over-the-shoulder on A. Camera behind B’s shoulder, looking at A. A speaks the important lines here. B’s shoulder and the back of their head sit soft in the foreground.
- Over-the-shoulder on B. The mirror. Camera behind A, looking at B, for B’s key lines.
- Reaction close-up on A. A listening. No dialogue. This is your editing glue and your emotion shot.
- Reaction close-up on B. Same job, other face.
That is the whole kit. Five generations, each short, each with a single responsibility. On a recent storyboard-first project I planned a twelve-line exchange across exactly these five setups, and the edit came together in about twenty minutes because every line had somewhere natural to live. Compare that to the old approach — one heroic 20-second wide, regenerated six times, still unusable — and the arithmetic isn’t close. Coverage costs a little more planning and a lot less re-rendering.
Keep the two OTS shots at matching distances. If A is framed chest-up and B in tight close-up, cutting between them feels like shouting across a canyon. Match the sizes, match the energy. Generate reaction shots longer than you think you need — a listener who holds for four usable seconds will save your edit twice.
Prompting One Speaker At A Time
The core prompting discipline for dialogue: in any given clip, one person talks and everybody else reacts. The moment a prompt asks two generated actors to take turns speaking inside a single clip, you’re asking the model to direct, act, and edit simultaneously. It will do all three badly.
So each shot’s prompt names a single job. For an OTS speaking shot, describe who is speaking, the emotional temperature in plain terms, and crucially what the listener’s shoulder is doing — ideally very little. Small prompt habits that pay off here:
- Keep the foreground shoulder simple. “The back of his head and shoulder, out of focus, still” — a busy foreground figure will pull detail and motion away from your speaker.
- Ask for modest mouth movement. If you are adding voice in post, which I recommend for anything beyond a few words, the actor mostly needs believable speaking rhythm, not precise phonemes. Slight head movement, natural pauses, eyes on the listener.
- Describe listening, not idling. For reactions, “listening, slight nod, expression shifting from neutral to concerned” gives the model an arc. “A man sits still” gives you a statue with blinking problems.
- Anchor every shot to the same references. Same character sheet, same wardrobe wording, same lighting description across all five setups. The reference image workflow matters more in dialogue than anywhere, because the audience stares at faces for the entire scene.
On voices: generating native synced dialogue inside the video tool works for short, punchy lines in tools that support it, and the results keep getting better. But for a full scene, recording or generating the voices separately and cutting picture to the audio gives you control the all-in-one route can’t. You can fix a line reading without regenerating a frame. You can tighten a pause by three frames. Most of my dialogue scenes are built picture-first with placeholder timing, then voiced, then re-trimmed to the actual performances. It sounds slower. It’s faster, because nothing sends you back to the render queue.
Eyelines And Screen Direction For Generated Actors
Here’s the unglamorous detail that separates a believable generated conversation from an uncanny one: where the actors are looking. In coverage, A must look toward B’s position and B must look back along the same axis. When eyelines match, the audience’s brain stitches two separately generated clips into one shared space. When they don’t — A gazing off left, B gazing off left too — both actors appear to be talking to an invisible third person standing off-screen. And the scene quietly falls apart.
Prompting eyelines takes a sentence of discipline per shot. Decide the geography once, on the storyboard: say A sits on the left side of the frame’s world, B on the right. Then in every OTS and close-up, A looks toward frame right, B looks toward frame left, at roughly matching heights — eyes slightly off lens, never into it. I write the eyeline into the prompt in plain words: “looking toward the right edge of frame, at eye level, as if listening to someone seated opposite.” Models handle this better than you’d expect when the instruction is concrete, and badly when it’s left to chance.
Screen direction matters just as much if anyone moves. If a character walks across the wide shot traveling left to right, cutting to them still traveling right to left in the next shot reads as them turning around and coming back. Generated clips make this mistake constantly because each one starts from a blank slate. Your storyboard is the only place the rule lives, so write it there: arrows for movement, an L or R next to each character’s name, and the eyeline direction for every close-up. Thirty seconds of notation saves a scene.
If a great take has the actor looking clean the wrong way, no expression rescues it. Re-generate. The audience won’t know why the scene feels wrong. They’ll just leave.
Cutting The Conversation In The Edit
Assembly in Premiere Pro is where coverage pays you back. My usual order of work: lay the wide down first for the opening seconds, then build the conversation from the OTS shots as the spine, and treat the reactions as a repair kit and a rhythm tool. A few cutting habits that make generated dialogue feel human:
- Cut on the listener, not just the speaker. If B lands a hurtful line, the interesting shot is A’s face receiving it. Cut to the reaction a few frames before you think you should; audiences read emotion off the listener.
- Let reactions cover weak mouths. Any stretch where lip movement looks mushy can play over the other actor’s reaction while the voice continues. This is the oldest trick in editing and it works on generated footage exactly as well as on filmed footage.
- Use J and L cuts constantly. Let the next speaker’s voice start a beat before the picture cuts (a J cut), or let a voice finish over the listener’s face (an L cut). Dialogue that only changes picture and sound on the same frame feels robotic even when every clip is technically fine, and this is the main cure.
- Tighten pauses ruthlessly. Generated clips are full of half-second dead air where nobody is quite acting. Steal those frames. Short-form dialogue should feel about 15 percent faster than feels natural in the timeline; it will feel right on a phone.
The fuller edit workflow — organizing generations, syncing separate voice tracks, where to put your trims — sits in the guide to editing AI clips in Premiere Pro, and it applies to dialogue scenes without modification. The one dialogue-specific addition: keep every reaction shot in a labeled bin and don’t delete the outtakes. In dialogue edits, yesterday’s rejected listener becomes the exact cover shot you need at midnight.
Sound deserves a sentence of its own. A faint room tone under the whole scene — even a generated ambience loop at low volume — hides the joins between clips more effectively than any visual trick, because each generation arrives with its own silent, subtly different acoustic fingerprint. Continuous sound tells the ear it’s one room and one moment. Cut the picture all you like; keep the room unbroken.
When A Voiceover Scene Beats A Talking Scene
Now the honest counterpoint, because coverage isn’t always the right call. Full dialogue coverage is the most generation-hungry way to tell a story with people in it, and sometimes the story doesn’t need faces talking at all.
Choose voiceover when the scene is really about memory, confession, or explanation — one voice telling us something while we watch it happen. A narrator over a character’s actions sidesteps lip sync entirely, halves your consistency risk, and often fits short-form better, because the voice can start on frame one with zero setup. Choose voiceover too when your characters are apart: a phone call rendered as two talking heads is fine, but a voice over what each person is doing while they talk is almost always more watchable.
Choose real coverage when the exchange itself is the event — an argument, a confession face to face, a negotiation, a joke that lives in the timing between two people. Interruption, overlap, and reaction are things voiceover can’t do. If the scene’s energy comes from two people affecting each other, pay for the five shots.
A hybrid I use often: voiceover carries the scene, and exactly one line is delivered on camera, in a close-up, at the turning point. The single synced line lands like a thunderclap precisely because everything around it was narrated. Audiences lean in when a character suddenly speaks for real. That trick alone is worth keeping in the drawer.
Whichever route you take, plan it on the board before you generate. Dialogue punishes improvisation more than any other scene type because the shots must agree with each other — eyelines, sizes, light, faces. Five planned generations and one honest decision about whether the scene even needs talking heads will outperform a dozen hopeful renders every time. That’s not a tool limitation. That’s just directing.