Voice-First AI Directing: Director Studio, from Scene to Screen
Learn voice-first AI directing with Director Studio: spoken commands, persistent staging, Cinematography Briefs, continuity checks, and reusable workflows.
TL;DR: Voice-first AI directing turns spoken camera and blocking instructions into changes to a persistent scene rig. Director Studio accepts voice or text, resolves the instruction, records the operation, and translates the resulting staging into a Cinematography Brief for generation. You still review the plan and the rendered take.
What voice-first AI directing actually means
Typing a cinematic prompt and directing a scene are different activities. A prompt describes an output. Direction changes a production decision: where the camera stands, whom it frames, how a performer turns, or what happens during the next beat. When every shot starts with a blank text box, those decisions must be restated, and contradictions become easy to miss.
Anamorpha's Director Studio connects natural-language direction to a persistent 3D scene rig. Say “over Sarah's shoulder on George” and the instruction has a spatial context: Sarah and George already occupy positions in the scene. The next cut can inherit that blocking rather than inventing another arrangement. Voice is the input method; persistent scene state is what makes the workflow useful.
This does not mean a microphone guarantees a better image than a keyboard. The console also accepts typed instructions. The advantage is working with addressable scene decisions instead of repeatedly rewriting an entire description. Use whichever input lets you express the next change most clearly.
Start with a scene, not a string of commands
Establish the dramatic question
Consider a short interrogation in a kitchen. Sarah wants George to admit he took a letter. George wants to leave without answering. The scene begins with both characters beside a table, and the letter remains between them. This gives the director a reason for each angle: the wide establishes distance, Sarah's close shot applies pressure, and George's reverse reveals avoidance.
Before moving a camera, prepare character references and a location reference. Decide where the table, doorway, and window sit. Name the performers consistently. “Move him closer” is less reliable than “move George toward the table” when multiple people appear. Spatial specificity is not extra paperwork; it prevents an ambiguous instruction from becoming an unwanted change.
Inspect the initial rig
The Director Console's top-down mini-viewport shows character positions, a camera glyph, and the scene axis. Check this starting arrangement against the story. Sarah should not accidentally stand between George and the only exit unless that obstruction is intentional. Confirm which way each character faces, because position and facing communicate different information.
The [persistent 3D scene rig guide](/blog/persistent-3d-scene-rig) explains this world state in detail. For a first exercise, keep the blocking simple: two stationary characters, one shared landmark, and a clear line of action. Add walking and moving cameras after the basic coverage cuts together.
How Director Studio resolves an instruction
Director Studio uses a three-tier resolver: workflow-memory replay, a local parser, and LLM reasoning through the director-voice-ops skill. The order matters. A previously resolved workflow can be replayed; recognizable commands can be handled locally; instructions requiring more interpretation can reach the reasoning layer. This is not the same as sending every utterance directly to a video model.
| Resolver stage | What it contributes | What the director checks | |---|---|---| | Workflow-memory replay | Reuses a saved resolved command | The current scene matches its assumptions | | Local parser | Resolves supported command patterns | The intended subject and operation are correct | | LLM reasoning | Interprets more complex direction | The interpretation preserves staging and intent |
The resolved operation is audited. That record helps answer a practical question: did the wrong result begin with an ambiguous instruction, a mistaken scene change, or the generated video? Separate those problems before revising. If the rig is wrong, generating more takes from it is unlikely to solve the underlying issue.
Push-to-talk uses the browser's Web Speech API. Speech availability and microphone permissions depend on the browser and device. If recognition is unavailable or a noisy room makes it unreliable, type the same direction. The production should not depend on having a quiet recording environment.
Direct the first shot in small, inspectable steps
Set framing before adding motion
Begin with an establishing view that includes both characters and the table. Then request an over-the-shoulder composition behind Sarah, framing George. Inspect which shoulder occupies the foreground and whether George's look direction still matches the wide. A correct name in the instruction is not enough; the spatial result must support the edit.
Next, set the focal length, for example 85mm. Treat that as a framing and perspective decision, not a universal “cinematic” switch. A longer lens may demand a different camera distance to maintain your chosen subject size. Check the composition again instead of assuming the lens change has preserved every visual relationship.
Finally, add a slow push-in. Specify the narrative purpose: George is running out of room to avoid the question. A restrained camera move can increase pressure without needing an exaggerated facial performance. Keeping framing, lens, and movement as separate decisions makes unexpected changes easier to identify.
Try this in Anamorpha: Open the [Director Studio overview](/product/director), then plan a two-person exchange. Give one framing instruction, inspect the rig, add one lens instruction, and inspect again before generating a take.
From the rig to a generation-ready brief
The Cinematography Brief translates spatial state into natural-language camera, staging, and motion instructions for the Seedance pipeline. An example is “eye-level 85mm OTS behind Sarah, framing George; Sarah stage-left facing right, George stage-right facing left.” The exact wording matters less than preserving the relationships that make the cut understandable.
Review the brief for contradictions with your shot description. If the rig says George faces Sarah but another sentence tells him to look through the window, clarify whether the look is a momentary action or a new facing. If a close shot should show only George, do not carry over a description demanding a balanced two-shot.
The brief guides generation; it is not a mathematical guarantee that every output follows the plan. Video models can still drift, misread a gesture, or distort a reference. Compare the generated take with the intended blocking, then decide whether to repair the source direction or rerender an otherwise sound shot.
Cut to the reverse without rebuilding the scene
For Sarah's reverse, change the camera while preserving the characters' established positions. That distinction is central: changing the viewpoint should not silently move the performers. The persistent rig provides a common spatial reference for the new angle, so the scene remains an arrangement you are photographing rather than a fresh illustration.
Watch the line-of-action warning before rendering. If the reverse places the camera across the established axis, you may reverse apparent screen direction. Either move back to the intended side or plan a motivated transition. The [180 degree rule guide](/blog/180-rule-ai-video) covers neutral shots, visible crossings, and deliberate exceptions.
Check eyelines as well as left and right. A seated character looking at a standing person should not stare horizontally merely because the horizontal axis is correct. Review the cut as a relationship between two views, not as two individually attractive compositions.
Save workflows after understanding their assumptions
DirectorWorkflow save-and-replay lets you reuse resolved commands. Useful candidates are recurring camera setups or a familiar sequence of changes. Save a workflow after verifying what it does, and give it a name that explains its purpose. “Quiet reaction close-up” is more informative than “good shot three.”
Replaying an operation does not remove the need to inspect context. Another scene may have different character spacing, room geometry, or dramatic intent. A camera move that worked in a large kitchen may be inappropriate in a corridor. Reuse a decision pattern, not the assumption that every scene is interchangeable.
Review performance separately from camera mechanics
A technically coherent scene can still feel empty. Once the coverage is stable, direct what the audience can observe. Instead of “George feels guilty,” ask for a delayed answer, a glance toward the letter, or fingers that stop tapping when Sarah names it. Keep the action small enough for the shot duration.
Character personality informs AI Director's performance planning, while motion notes specify the physical beat. Read [character subtext in AI video](/blog/character-subtext-ai-video) for a method that separates inner motive from visible behavior. Camera direction and acting direction should reinforce the same dramatic event, not compete for attention.
A practical completion checklist
- Confirm the named characters, references, positions, and facings before the first take.
- Review the resolved operation after each meaningful camera or blocking change.
- Check the axis and eyelines when moving between coverage angles.
- Read the Cinematography Brief alongside the intended action.
- Judge generated takes in sequence, including entry and exit poses.
- Save reusable workflows only after documenting their scene assumptions.
Voice-first AI directing is most effective when it shortens the path from intention to an inspectable production change. Speak naturally, make one important decision at a time, and keep the rendered sequence as the final test. [Explore Director Studio](/product/director) to start with a scene that remembers your staging.
Continue exploring
- AI Scene Continuity: Inside the Persistent 3D Scene Rig
- 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