AI Scene Continuity: Inside the Persistent 3D Scene Rig
Build AI scene continuity with Scene3D: persistent character positions, camera poses, facings, and staging that carries from one shot to the next.
TL;DR: AI scene continuity needs more than matching faces. Anamorpha's Scene3D keeps character positions, facings, camera pose, and the line of action in a persistent rig. Each shot can inherit that staging, and a Cinematography Brief carries it into the video prompt. The rig supports continuity planning; rendered footage still needs review.
Why convincing shots can make an incoherent scene
A woman stands beside a doorway in the wide. In the reverse, the doorway appears behind the man. In the close-up, she looks the opposite way. Each image may be polished, but the sequence no longer describes a room the audience can understand. This is a spatial continuity problem, not simply a failure to match wardrobe or facial identity.
Independent prompts encourage this failure because they describe each shot without a shared coordinate system. “Sarah on the left” might mean the left of the image, the left side of the set, or Sarah's own left. Move the camera and those meanings diverge. A persistent 3D scene rig gives the production a world-space reference that survives changes in viewpoint.
AI scene continuity therefore has several layers. Identity references help preserve who is present. Set references establish what surrounds them. The rig records where subjects and the camera are placed. Performance notes describe what happens over time. None of these layers replaces all the others.
What Scene3D remembers
Positions and facings are separate decisions
A character's position tells you where they stand; their facing tells you where their body points. A person can remain beside a table while turning toward the door. Treating that turn as a new location would create an unnecessary blocking change. Treating a walk as merely a turn would miss a real change in spatial relationships.
Scene3D persists both positions and facings, alongside the camera pose. That allows a new angle to use the existing arrangement rather than infer it again from prose. Directors can inspect these relationships in the top-down mini-viewport, where a simple layout is often easier to judge than a heavily styled reference frame.
The line of action is part of the scene
The sceneAxis records the line of action used to reason about continuity. For a two-person conversation, this usually relates to the line connecting the participants. Cameras placed consistently on one side help preserve apparent screen direction across cuts. When blocking changes, the director must consider whether the established axis still describes the action.
This is a production world model, not a claim that the video generator runs a complete physical simulation of the set. The value is persistent staging information that can be inspected, refined, and translated into instructions. It reduces repeated interpretation without making every downstream frame deterministic.
Build a small set that supports the scene
Start with three meaningful landmarks rather than dozens of decorative objects. In a kitchen conversation, use the table, the window, and the exit. Decide which landmark anchors each character. Sarah stands at the window end of the table; George remains nearer the door. Now a request to “move George toward the table” has a useful reference.
Location Studio supports semantic 3D layout extraction from set images, including depth and floor-plan information used for blocking. Treat the extracted layout as something to inspect, especially when a reference has unusual perspective, mirrors, or partially hidden surfaces. A visually convincing image does not automatically yield a perfectly understood walkable space.
Keep scales and distances plausible for the intended coverage. If characters stand too close, an over-the-shoulder shot may obscure the subject. If they stand very far apart, an intimate reverse may require an implausible camera position. Resolve those constraints in the staging plan before paying for motion generations.
One scene, three shots, shared blocking
Shot one: establish the geography
Frame Sarah, George, the table, and enough of the exit to explain the room. The audience should understand that George can leave and Sarah is not physically preventing him. This establishes the dramatic relationship as well as the spatial arrangement. Record the starting facings before planning tighter coverage.
Shot two: change the camera, not the actors
Move to an over-the-shoulder angle behind Sarah on George. Preserve both characters' positions unless the script calls for movement. Compare the camera side with the wide. George's glance toward the door should still point toward the same destination, even though the door may now sit outside the frame.
Shot three: give movement a cause
After Sarah mentions the letter, George steps closer to the table. This is a new blocking beat, not a correction to the previous shot. Update the intended scene state for that beat, then plan the next angle from the resulting arrangement. Distinguishing an intentional move from an accidental reset protects both story and continuity.
Try this in Anamorpha: Use [Director Studio](/product/director) to plan a wide, an over-the-shoulder shot, and a reverse around one table. Inspect the top-down view after every change and ask whether you moved the camera, a character, or both.
How the Cinematography Brief connects planning to pixels
The rig is translated into natural-language camera, staging, and motion guidance. This Cinematography Brief is injected into the Seedance video prompt, so the generation request carries the arrangement you planned. Instead of relying solely on “tense conversation,” the prompt can specify who faces whom and how the camera views that relationship.
For example, a brief might describe an eye-level view behind Sarah with George framed across the table. The performance instruction can then add that George pauses before reaching for the letter. Keep these responsibilities clear: the rig supplies spatial constraints; the dramatic direction supplies meaningful action within them.
| Continuity layer | Main question | Useful source | |---|---|---| | Character identity | Is this the same person? | Character sheet and selected variant | | Spatial staging | Where is everyone? | Persistent Scene3D rig | | Shot composition | From where do we see the action? | Camera pose and lens direction | | Performance | What changes during the beat? | Subtext and motion notes | | Final continuity | Does the cut actually work? | Rendered clips reviewed together |
Avoid competing instructions. A location reference showing a doorway behind Sarah and a text prompt placing it behind George create unnecessary uncertainty. Fix the disagreement at the source rather than adding increasingly emphatic prose to the same contradictory request.
What happens when a new shot conflicts with staging
Some requested angles do not fit the existing arrangement cleanly. You may want a reverse that crosses the axis, an eyeline that misses the partner, or a camera position that makes the blocking unreadable. The Scene-Pose Solver is an escalation skill for refining the rig when a shot conflicts with prior staging, with screen direction and eyeline preservation as goals.
The director still decides which constraint matters most. Perhaps a character must move to motivate the composition. Perhaps the shot should become a single rather than an over-the-shoulder. Perhaps a wider transition is needed. A solver can help resolve spatial conflicts; it cannot decide the dramatic meaning of every compromise for you.
When reviewing a proposed adjustment, compare it with the last approved beat. Do not evaluate only the new camera view. A composition that looks better in isolation may require a performer to jump across the room between cuts. Continuity depends on the transition, not merely the endpoint.
Persistent state does not eliminate time
A scene rig is a spatial reference, while a take unfolds over time. A hand moves, a chair turns, a character crosses the frame. Review the end of one clip against the beginning of the next. Even if both shots started from sensible positions, their motion may leave them in incompatible states at the planned cut.
For actions involving a prop, choose an editorial handoff: before the reach, during contact, or after the object is lifted. Record which hand holds it and where it ends. These practical notes complement the rig instead of asking it to represent every detail of a performance.
Moving cameras require the same care. A starting camera pose may be on the correct side of the axis while the planned travel crosses it. Read the [180 degree rule guide](/blog/180-rule-ai-video) before designing an orbit or a moving reverse. Inspect the intended path and the rendered movement, not just a single top-down snapshot.
A review method for AI scene continuity
Review a short assembly without sound first. Can you tell who is looking at whom? Does movement continue toward the same destination? Are exits and important objects where the establishing shot suggested? Silent playback makes spatial problems harder to hide behind compelling dialogue or music.
Then restore audio and assess dramatic rhythm. A technically matched cut can still arrive too early for a reaction to register. Let the scene's intention determine the final edit while keeping the geography legible. Technical coherence is a foundation for storytelling, not the complete story.
For recurring locations in a longer production, pair scene-level staging with series-level references. The [AI film series workflow](/blog/ai-film-series-workflow) explains how a shared bible and inherited style support episodes without making every scene identical. Reuse the world's rules while allowing characters to occupy it differently.
Build continuity before generating coverage
The most useful habit is to separate world decisions from viewpoint decisions. Establish the space, place the performers, choose an axis, and then explore cameras. When the story changes the blocking, make that change explicit. When only the camera moves, preserve the rest.
Read the [voice-first directing walkthrough](/blog/voice-first-ai-directing) for an end-to-end command workflow, or [explore Director Studio](/product/director) to begin planning coverage. A persistent rig helps your shots describe one scene instead of several attractive but incompatible versions of it.
Continue exploring
- Voice-First AI Directing: Director Studio, from Scene to Screen
- The 180 Degree Rule in AI Video: Catch Axis Crossings Before Render
- Character Subtext in AI Video: Render the Tell, Not the Emotion Label
- An AI Film Series Workflow: Containers, Shared Bible, and Inherited Style