How to Generate Eating and Drinking Scenes in AI Vertical Drama So Food and Glasses Stay Consistent
A family dinner. The reveal happens over the main course. In the wide, the glass in front of the heroine is half full of red wine. In her close up it is full. In the reverse it is empty, and her plate has changed from steak to pasta. The dialogue is strong and the performance lands, but the comments under the episode are about the wine. Meal scenes are where AI vertical drama continuity breaks most often, and where viewers notice it most quickly.
The reason is simple. A table holds dozens of small objects, each of which changes state as the scene goes on. Plates empty, glasses drain and refill, cutlery moves, napkins fall. In a live action shoot a props team resets the table between takes and a script supervisor photographs every state. In a generated scene, every shot rebuilds the table from instructions, and the generator has no memory of how much wine was in the glass two shots ago. The fix is to treat the table as a location, the food and drink as tracked props, and consumption as a state that the production records rather than an action the generator improvises.
1. Why Meal Scenes Break Faster Than Any Other Scene Type
Most continuity problems involve one or two elements. A face, a costume, a prop. A meal scene multiplies them. A dinner for four can easily carry four plates, four glasses, a bottle, a serving dish, bread, cutlery for each setting and a centrepiece. Each object has a position, an appearance and, for food and drink, a quantity. Every shot in the scene is a fresh chance for any of those to change.
Quantity is the hardest part. Faces and costumes are meant to stay the same. Food and drink are meant to change, but only in one direction and only at a believable rate. A glass that goes from half full to full with no one pouring is an error. So is a plate that empties in one cut. The generator cannot track this on its own, because it sees each shot in isolation. It will render whatever level of wine looks plausible for a dinner table, and plausible changes from generation to generation.
Liquids add a technical layer. Generators often struggle with transparent glass, reflections and the surface of a drink. Wine can turn opaque, water can vanish, and the glass itself can change shape between a tulip and a tumbler. Hands make it harder still, because a hand around a glass hides part of it and the model has to guess what is behind the fingers.
Vertical framing makes all of this more visible. Meal scenes in 9:16 are usually covered with close singles across the table, and a single often includes the glass and the edge of the plate in the lower part of the frame. The objects sit right where the eye rests between lines of dialogue.
2. Decide What the Meal Is Doing in the Story
Before any table is designed, the production should answer a simple question. What is the meal for? In vertical drama a dinner scene is rarely about the food. It is a pressure cooker: people who cannot easily leave, forced into one room, with an excuse to sit close and watch each other react. The meal is a setting for conflict, not the subject.
That answer shapes how much eating and drinking the scene actually needs. If the point is a confrontation, the characters may barely touch the food. A glass of wine can carry the tension on its own, raised, set down, pushed away. If the point is a seduction, the shared drink may matter more than any plate. If the point is a family showing a perfect surface before it cracks, the table may be beautifully laid and almost untouched.
The less on screen eating a scene requires, the easier it is to keep consistent. Every bite and every sip is a state change that must be tracked. A script that has characters eat continuously while delivering long speeches creates far more continuity work than one where the meal sits in the background and only a few deliberate actions carry meaning. Writers and directors working on AI vertical drama benefit from treating each eating or drinking action as a choice that has to earn its place.
Mark the meaningful actions in the script
Once the scene purpose is clear, the script should mark every food or drink action that matters. A glass raised for a toast. A drink thrown. A fork set down hard when the secret comes out. A plate pushed away. These are the beats the generation team will protect most carefully. Everything else at the table can sit still, which is the easiest state to hold consistent.
3. Build a Table Plan Before Generation
Every meal scene that runs longer than a few shots needs a table plan. It is a simple top down drawing of the table, showing where each character sits, where each plate, glass and serving dish is placed, and which side of each setting the cutlery and glass sit on. It takes a few minutes to draw and saves many retakes.
The table plan also fixes the geography of the scene. Who sits on which side, who faces the camera in which shots, and which way each character looks to reach the others. This connects directly to eye line work. A dinner table is one of the most complex dialogue setups in drama, because several people speak to each other across and along the table, and every look has to point the right way.
Alongside the plan sits a short table specification. The style of plates and glasses, the cloth or bare wood, the type of cutlery, the centrepiece, the lighting. Like any location record, it should be written once and pulled verbatim into every prompt, so that two operators generating different shots describe the same table in the same words.
Treat the table as a location in the library
For a recurring dinner table, such as a family kitchen that appears in many episodes, the table belongs in the location library. Generate a set of empty table reference frames from the main camera angles, reviewed together so that the geometry and light agree, and lock them. Scenes at that table then generate against those frames rather than describing the table from scratch. The same principle that keeps a living room stable across a long series keeps a dinner table stable across a long scene.
4. Track Consumption as State, Not Action
The most effective single change a team can make is to stop asking the generator to show eating as an ongoing process and start defining consumption as a series of fixed states.
A state is a snapshot. Plate full, plate half eaten, plate cleared. Glass full, glass half, glass empty. Bottle full, bottle half, bottle on its side. Each state has a short locked description and, for hero items, a reference image. The scene is then divided into segments, and each segment has one defined state for every item on the table.
The state register for a scene is a short table of beats against items. Beat one, before the reveal: all plates full, all glasses half full. Beat two, after the reveal: her glass empty, his glass untouched, her plate pushed aside. Beat three, the exit: her chair empty, her napkin on the table. Every shot in the scene belongs to one beat, and every prompt in that beat carries the same state lines.
This turns an impossible task into a manageable one. The generator is never asked to show wine level falling smoothly across a scene. It is asked to show a glass at half, then a glass empty, with the change happening at a cut or in a single deliberate action. The audience accepts this naturally, because cuts in any drama compress time. What they do not accept is a level that goes up without anyone pouring.
Change state at the cut or on a single action
State changes should happen in one of two places. At a cut, when time is understood to have passed, or inside one clearly framed action such as a character draining a glass. Changing state mid shot without a visible cause invites the generator to interpolate, and interpolation is where levels and contents go wrong.
5. Keep Mouths and Hands Out of Trouble
Eating on camera is difficult for generators. Chewing, swallowing and lifting food to the mouth involve fine movements of the jaw, lips and hands, often all at once. The results can look rubbery or unsettling, especially in a close up. Drinking is slightly easier but still involves a glass partly hiding the face, which can disturb the identity of the character.
The first rule is to avoid eating while speaking. A character chewing through a line of dialogue is hard to generate convincingly and harder to lip sync. Direct performers, in the script, to finish a mouthful before speaking, or to set the fork down and speak, or to speak first and then eat. This is natural behaviour at a tense table anyway, and it gives every line a clean, readable face.
The second rule is to keep bites small and gestures simple. A fork lifted to the mouth and lowered again is easier to hold than a knife and fork cutting meat in close detail. A sip is easier than a long drink. Where the story needs a bigger gesture, such as a character downing a glass in one go, break it into short clips: the hand lifting, the drink at the lips, the glass set down empty. Each clip has its own state, and the edit carries the action.
Watch the hand around the glass
A hand wrapped around a glass is a common failure point. The fingers can merge with the stem, the glass can change shape behind the hand, or the liquid can vanish. Keeping the grip simple, with the glass clearly visible and the liquid level readable, makes a large difference. The glass should be a locked prop with its own reference, including an in hand image at the scale it appears on screen.
6. Pour, Toast, Spill: Handling Liquid Action
Some meal scenes depend on a liquid action. A toast at a wedding. A glass of wine thrown in a face. A drink spilled across a contract. These moments are often the visual peak of the scene, and they are among the hardest things to generate well.
Pouring works best when it is short and framed tightly. A bottle tilting, a stream entering the glass, the level rising to a clear final state. Generate it as a dedicated insert, approve it, and use a frame from the end of the pour as the reference for the shots that follow, so the glass level in the next shots matches what the audience just saw.
A toast is mostly about faces and raised arms, with the glasses at the top of the frame. Keep the glasses at a consistent level across the singles in the toast, and treat the clink, if the story needs it, as a separate insert. Sound does a great deal of the work. A clean clink on the audio track sells the moment even if the glasses meet only briefly on screen.
Throwing and spilling are the most demanding. Liquid in motion is unpredictable for generators and can turn into smoke, glass or noise. The most reliable approach is to split the action into before, during and after. The arm moving, a short burst of liquid, and then the result: a wet face, a stained shirt, a soaked page. The result shot carries the story, so it deserves the most generation effort and the strictest review. The aftermath must then stay consistent for the rest of the scene, which means a new state for the wet shirt and the empty glass, logged in the register.
7. Review the Table, Not Just the Faces
General quality review in AI vertical drama focuses on faces, hands, motion and light. In a meal scene, that review will pass shots where the wine level jumps and the plates swap, because each shot looks fine on its own. The errors only exist between shots.
A meal scene needs a table pass. The reviewer lays the shots in order and watches the scene as a viewer would, then checks each shot against the state register. Is every plate, glass and bottle in the correct state for the beat this shot belongs to? Is every object in the position the table plan says it should be? Is the glass the same glass? Is the cutlery on the correct side?
The same attention to small quantities is a classic concern of continuity in conventional film, where the level of a drink between shots is one of the most familiar examples of a visible error. In a generated scene the risk is higher, because nothing physical holds the level in place between shots, so the check has to be deliberate.
Write retake notes in state terms
When a shot fails the table pass, the retake note should name the shot, the item, the wrong state and the correct state from the register. Shot six, her glass, reads full, should be empty per beat two. A note in that form lets any operator correct the shot in one pass, and the accumulated notes show which items fail most often so the team can strengthen their references and descriptions.
Use the edit to protect the scene
Some table errors are small enough that a retake is not worth losing a strong performance. In those cases the edit has options. A slight reframe can move the glass out of the frame. A cut to a reaction or an insert can bridge two shots with a small mismatch. A close up that keeps the table below frame can carry a line that a wider shot cannot. These are rescue tools, not plans, and a scene that needs many of them was not tracked closely enough at the start.
Axis AI Studios Perspective
Meal scenes look simple and behave like some of the most complex scenes in a series. A table full of objects changing state, several characters speaking across it, and hands constantly moving between plate, glass and face. In AI vertical drama the generator holds none of that in memory, so the production has to hold it instead.
Axis AI Studios treats a meal as a location with tracked props and defined states. Tables are planned and referenced like rooms. Consumption is broken into beats with fixed states, changed only at cuts or in single deliberate actions. Eating is kept off spoken lines. Liquid actions are split into short clips with clear before and after states. And every meal scene gets a table pass against its register before it is approved. The result is dinner scenes where the conversation carries the drama and nobody in the comments is counting glasses.
For platforms and producers with scripts built around family dinners, celebrations, business lunches or any scene where the table matters, write to business@axisaistudios.com to talk through how the production would handle it.
FAQ
Why do wine glasses and plates change between shots in AI generated meal scenes?
Each shot is generated separately, and the generator has no memory of the quantity or position of objects in the previous shot. It renders whatever looks plausible for a dinner table, and plausible varies between generations. Defining consumption as fixed states per beat, with locked descriptions and references for hero items, removes most of this variation.
Should characters eat while they deliver dialogue in AI vertical drama?
Generally no. Chewing while speaking is difficult to generate convincingly and harder to lip sync, and it often disturbs the face in a close up. Directing characters to finish a bite before speaking, or to set the cutlery down and speak, produces cleaner shots and suits the tension of most vertical drama dinner scenes.
How should a production handle a drink being thrown or spilled?
Split the action into short clips: the movement, a brief moment of liquid in motion, and the result. The result shot, such as a wet face or a stained document, carries the story and deserves the most effort and review. From that point on, the aftermath becomes a new state in the register and must stay consistent for the rest of the scene.
Further Reading
For teams applying the same discipline to every object a character holds, the guide to keeping props and hand held objects consistent across shots covers prop classes, reference sheets and how inserts anchor an object for the audience.
For directors staging a conversation across a dinner table, the method for matching eye lines across separately generated shots covers scene maps, gaze vocabulary and reviewing a dialogue scene as a cut.
For production leads tracking state changes across a full run of episodes, the walkthrough on building a vertical drama continuity bible covers how character, environment and object states are recorded so nothing drifts.

Let's set
the new standard together.
If you're working on something, we'd like to hear about it.
