How to Match Eye Lines Across Separately Generated Shots in AI Vertical Drama Dialogue Scenes

Two shots. Two clean faces. One broken scene. The editor lays the single of the heroine against the single of the man she is confronting, and something is wrong before anyone can name it. She looks up and to the left. He looks straight into the lens. The viewer feels the two characters are not in the same room, and the argument that should land as the emotional peak of the episode reads as two people talking to someone off screen.

This is the most common reason a dialogue scene in AI vertical drama fails review even when every individual shot passes. Each shot is generated in its own session, from its own prompt, often by a different operator on a different day. Nothing in the generator knows where the other character is standing. Eye lines are a relationship between shots, and the generator only ever sees one shot at a time. The fix is not a better prompt for any single shot. It is a shared map of the scene that every shot is generated against, and a review step that checks the relationship rather than the frame.

1. Why Eye Lines Break More Often in Generated Dialogue

In a live action shoot, eye lines are mostly free. The actors stand where they stand, the camera moves around them, and the geometry of the room enforces itself. The camera operator can see the other actor, or a mark on a stand, and the performer looks at it. Mistakes still happen, but the physical world does most of the work.

Generation removes that physical anchor. Every shot is a fresh composition, and the model decides where the eyes point based on the words in the prompt, the reference frame, and its own training habits. Most models have a strong default toward a near lens gaze in close ups, because so much of the material they learned from is portrait and social content. Ask for a close up of a woman listening and you will often get a woman listening to the viewer.

The vertical frame makes this worse. In 9:16 there is very little horizontal room, so the small lateral shift that tells the viewer where the other character stands is compressed into a few degrees. A gaze that is slightly too centred reads as a look to camera. A gaze that is slightly too wide reads as a look past the other person. The margin for error in a vertical single is far narrower than in a widescreen single, and generated shots rarely land inside it by accident.

There is also a height problem. Characters in vertical drama are often staged with status built into position. One stands, one sits. One is on the stairs, one is at the bottom. If the seated character looks up in her single but the standing character looks straight ahead in his, the power dynamic collapses. Viewers read this instantly, even if they could not explain it.

2. Build the Scene Map Before Any Shot Is Generated

The core tool is a simple scene map, prepared by the director or the prompt writer before generation starts, and attached to every shot request in the scene. It is not a storyboard. It is a top down plan of where every character is, where the camera is for each shot, and which way each character is looking.

2.1 What the Scene Map Contains

At minimum the map records the position of each character in the room, the line of action between the two speakers, the side of the line the camera stays on, and the camera position for each planned shot. It also records the height relationship: who is higher, who is lower, and by roughly how much. For a two hander this fits on a single page and takes a few minutes to draw.

From the map, every shot gets a gaze instruction written in screen terms rather than story terms. Not that she looks at him, but that she looks frame left, slightly above lens height, at a point just outside the frame edge. Story terms mean nothing to a generator that has never seen him. Screen terms describe what the frame should actually contain.

2.2 The Line of Action

The classic continuity rule still applies. If the camera stays on one side of the line between the two characters, one character will always look frame left and the other will always look frame right, and the cut between them will feel natural. A useful primer on the principle is the 180 degree rule, which every operator on a dialogue scene should understand in practical rather than theoretical terms.

In generated production, the line has to be stated explicitly in every prompt, because the model has no memory of which side the previous shot was on. We write the screen direction of each character into the shot header as a fixed field, so the operator cannot forget it and the reviewer can check it against the frame without opening the map.

2.3 Height and Distance

Direction is only one third of an eye line. Height is the second, and distance is the third. A character looking at someone two metres away has a different eye convergence and a slightly different head angle from a character looking at someone half a metre away. In close ups, viewers pick this up from the eyes themselves.

The map records a rough distance for each pair, and the gaze instruction carries it through. Close, conversational, or across the room. It sounds like a small thing. It is the difference between a confrontation that feels intimate and one that feels staged.

3. Writing Gaze Instructions a Generator Can Follow

Even with a clean map, the prompt has to translate it into language the model responds to reliably. This is where most teams lose consistency, because different operators describe the same gaze in different words and get different results.

3.1 Use a Fixed Gaze Vocabulary

We keep a short controlled vocabulary for gaze, and every prompt in a dialogue scene uses it. Direction is always one of a small set of terms tied to frame edges. Height is always one of three or four terms tied to lens height. Distance is always one of three terms. The point is not that these exact words are magic. The point is that the same words produce the same behaviour across operators, and when they do not, the failure is easy to diagnose.

Free description is the enemy here. One operator writes glancing away, another writes looking off to the side, another writes eyes averted. The model interprets each differently, and the scene drifts. A fixed vocabulary removes that variation at the source.

3.2 Anchor the Gaze to Something in Frame

Models follow gaze instructions more reliably when there is something in the frame that implies where the other person is. A shoulder edge in the foreground. A soft shape at the frame border. A reflection. In an over the shoulder shot this is built in. In a clean single it has to be suggested.

When a clean single keeps drifting toward lens, adding a foreground element on the correct side often pulls the gaze across without any change to the face prompt. The element can be softened or cropped in the edit. Its job is to tell the model where to look.

3.3 Avoid Words That Pull Toward Lens

Certain words reliably pull the gaze toward camera. Anything that describes the character addressing, confessing, pleading or confronting without naming a direction tends to be read as addressing the viewer. These are exactly the verbs dialogue scenes are full of. The fix is to keep the emotional verb but always pair it with an explicit screen direction, so the model knows the confession is going frame right, not into the lens.

4. Reference Frames and the Anchor Shot

The most effective single technique for eye line consistency is to generate one anchor shot first and use it as the reference for the rest of the scene.

4.1 Generating the Anchor

The anchor is usually the widest shot in the scene that still shows both characters clearly enough to read their gaze. A two shot, or a loose over the shoulder. It establishes the geography: who is on which side, who is higher, how far apart they are. It is reviewed and approved before any single is generated, and its approval is specifically an approval of the eye lines, not just the faces and the light.

4.2 Deriving the Singles From the Anchor

Once the anchor is locked, each single is generated with the anchor as a reference where the tool supports it, and with a gaze instruction that matches what the anchor shows. If the heroine looks frame left and slightly up in the anchor, her single carries that exact instruction. If the tool supports image references for composition, a crop of the anchor that shows the direction of her gaze is often more effective than any amount of text.

For tools that allow camera or pose control, structural guidance can hold the head angle steady across the singles. This is especially useful when the same character appears in several singles across a long scene and the head keeps rotating slightly between them.

5. Reviewing Eye Lines as a Pair, Not a Frame

Most quality review checks one shot at a time. Faces, hands, light, artefacts. That review will pass every shot in a scene with broken eye lines, because no single shot is wrong. The error only exists in the cut.

5.1 The Cut Test

Every dialogue scene gets a cut test before it is marked complete. The reviewer lays the shots in their planned order, with rough timing, and watches the scene as a viewer would. Not frame by frame. As a scene. The question is simple. Do these two people feel like they are talking to each other in the same room.

If the answer is no, the reviewer identifies which shot breaks the relationship. Usually it is obvious once the scene plays. One gaze is too central, one is on the wrong side, or one is at the wrong height.

5.2 Checking Against the Map

After the cut test, the reviewer checks each shot against the scene map. Is the screen direction correct. Is the height correct relative to the other character. Is the distance consistent. This catches the subtler failures that the cut test might let through on a first pass, such as a gaze that is on the correct side but at a height that undercuts the power dynamic.

5.3 Writing the Retake Note

When a shot fails, the retake note names the specific eye line field that is wrong and the correct value from the map. Not that the gaze feels off. That her gaze is too close to lens, and the correct direction is frame left, just outside the frame edge, slightly above lens height. A note written this way can be acted on by any operator without a conversation.

6. Fixing Eye Lines in the Edit When a Retake Is Not Worth It

Not every broken eye line justifies a regeneration. Sometimes the performance in the flawed shot is the best in the set, and a retake risks losing it. In those cases, the edit has a few tools.

6.1 Reframing

A slight horizontal reframe can shift where the eyes sit relative to the frame centre, which changes how the viewer reads the direction of the gaze. In vertical, there is not much room to do this, but even a small shift can move a gaze from reading as a look to camera to reading as a look just past it. It works best when the original generation was done with some headroom and side room to spare, which is a reason to generate singles slightly looser than the final crop.

6.2 Cover With an Insert

A hand, an object, a reaction from a third character. A short insert between two singles with slightly mismatched eye lines gives the viewer time to reset, and the mismatch becomes much less noticeable. This is a rescue technique, not a plan. If a scene needs inserts to hide its eye lines, the scene map was not followed.

7. Making Eye Line Consistency a Pipeline Habit

The method above works on one scene. Across a 70 episode series, with dozens of dialogue scenes and several operators, it only works if it is built into the pipeline rather than left to individual care.

7.1 Scene Maps as a Required Asset

No dialogue scene enters generation without a scene map. The map is stored with the scene, versioned like any other reference asset, and linked from every shot request. If a scene is regenerated weeks later, the map is still there and the new shots match the old ones.

7.2 Gaze Fields in the Shot Record

Every shot record carries its gaze direction, height and distance as structured fields, not as free text buried in the prompt. This makes it possible to search for every shot where a character looks frame left, to check consistency across episodes, and to spot an operator who is repeatedly drifting from the map.

7.3 Recurring Locations, Recurring Geography

When a series returns to the same location for many scenes, such as the office, the kitchen, the hospital corridor, the geography of that location should be stable too. If the heroine always sits on the left side of the desk, her gaze across the desk is always frame right. A location library that records standard staging positions, alongside the visual reference, saves the director from redrawing the same map every episode and keeps the series feeling like one consistent world.

Axis AI Studios Perspective

Eye lines are a small detail that carries a large share of whether a dialogue scene works. In vertical drama, dialogue scenes are where most of the emotional weight of an episode sits. Confrontations, confessions, reversals. If the viewer does not believe the two characters are looking at each other, none of that weight lands, and no amount of polish on the individual faces will recover it.

At Axis AI Studios, eye lines are treated as a scene level asset rather than a shot level outcome. Every dialogue scene is mapped before generation. Every shot carries its gaze as a structured field. Every scene is reviewed as a cut, not as a set of frames. This is part of a wider principle in how we run production: continuity is planned, recorded and checked, never left to the generator to remember.

For platforms and commissioning teams, the practical outcome is dialogue scenes that hold together on first review, retake notes that are specific enough to act on, and series that feel like they were shot in one consistent world from the first episode to the last. If you are commissioning AI vertical drama and want to talk through how continuity is handled across a full series, write to us at business@axisaistudios.com.

FAQ

Why do AI generated dialogue shots often look like the character is talking to the camera?

Most video models have a strong default toward a near lens gaze in close ups, because much of their training material is portrait and social content that addresses the viewer. Without an explicit screen direction in the prompt, and ideally a reference or foreground element that implies where the other character is, the model tends to fall back on that default. Emotional verbs such as pleading or confessing make this worse unless they are paired with a stated direction.

Is the 180 degree rule still relevant when every shot is generated separately?

It is more relevant, not less. In a physical shoot the room enforces geography for you. In generated production nothing does, so the line of action has to be stated in every shot request and checked in review. Keeping the camera on one side of the line is the simplest way to guarantee that the two characters look in opposite screen directions and cut together naturally.

Can broken eye lines be fixed in post instead of regenerating?

Sometimes. A small reframe or a short insert can rescue a shot that is close to right. A shot that is badly wrong usually needs a retake, because heavy correction tends to make the eyes look artificial in a close up. The reliable fix is upstream: a scene map, a fixed gaze vocabulary, and an anchor shot that every single is generated against.

Further Reading

For anyone working through the face consistency side of the same problem, the guide to generating two character scenes without losing either face covers how to keep both identities stable when they share a frame.

For directors who want to plan the coverage that a scene map is built from, the walkthrough on storyboarding for the 9:16 frame covers how to plan shots and angles before production begins.

For operators who need to hold head angle and camera position steady across a run of singles, the guide to using ControlNet for consistent camera angles covers structural guidance for repeatable framing.