Naming the move, one move per shot, what models do with compound moves: camera mechanics, dialogue constraints, and previz rules.
OMNI Labs
Key takeaways
Every shot specification must mandate one move defining start position, spatial path, end position, and camera velocity.
Prompts must write 'dolly in' and explicitly append 'No zoom' to prevent diffusion models from substituting digital scaling.
During spoken dialogue on Seedance 2.5, only scale moves (slow dolly in/out, restrained handheld, locked-off) are allowed.
In Blender previz, dolly back with zoom in compressed background geometry by 2.4x; close-up dialogue framing uses ~29 degrees.
A silent 4-second 720p signature take costs $1.85 ($0.4622/s); 480p probe costs $0.82, and 4 s previz adds $2.22 total.
Single-move discipline and the essential 'No zoom' rule
Every shot in our plans carries one chosen camera move, written as start position, path, end position and speed. For identity-critical clips of 5 to 8 seconds we allow ONE main move; at 10 to 15 seconds, at most two compatible moves.
In generative video cinematography, camera trajectory directs narrative attention and establishes physical scale. However, compounding multiple camera instructions—such as demanding a simultaneous crane tilt, truck left, and orbital spin—confuses latent velocity fields, causing severe visual tearing and actor deformation. Studio prompt architecture enforces strict single-move discipline: every shot declares one move defined by initial camera position, spatial vector path, terminal frame position, and travel velocity. For identity-critical takes spanning 5 to 8 seconds, prompts permit exactly one primary move; takes extending from 10 to 15 seconds permit at most two mechanically compatible moves.
We write 'dolly in' for a camera that travels forward and never 'zoom' for a travelling move. Our paste-ready dolly line ends 'No zoom', because without that line the model may just scale the image.
Semantic precision governs optical behavior. In physical filmmaking, dollies translate the camera physically through space, altering perspective parallax, while zooms merely adjust focal length. Because video diffusion backbones confuse optical zooming with physical tracking, prompts must specify 'dolly in' when forward travel is intended. Crucially, studio templates append the explicit negative constraint 'No zoom' to the dolly instruction; omitting this rule frequently causes the model to crop and scale the image digitally rather than rendering genuine physical camera movement.
Dialogue camera constraints and narrative shot variety
While a mouth speaks on a Seedance 2.5 shot we allow only scale-only moves: slow dolly in, slow dolly out, restrained handheld, locked-off. Pans, trucks, tilts and arcs must finish before the line starts or begin after it ends.
Synchronizing synthetic vocal tracks with dynamic physical motion requires strict spatial boundaries. When talent delivers dialogue in ByteDance Seedance 2.5, lateral or rotational camera motion destabilizes facial landmark tracking, causing lip-sync de-synchronization. Consequently, during active speech, our directors permit only scale-preserving moves: gentle dolly in, gentle dolly out, subtle handheld drift, or locked-off framing. Rotational pans, lateral trucks, vertical tilts, and orbital arcs must conclude prior to speech onset or initiate only after dialogue terminates.
Our first-season draft came out flat: 47 of 48 shots were eye-level MCU or CU singles. The fix was a deliberate move on every shot, dialogue shots included, not more effects.
Early production reviews highlight the danger of static staging. In an unoptimized first-season episode draft, 47 of 48 shots defaulted to static eye-level medium close-ups or close-up single takes, creating a visually monotonous sequence. The solution required designing deliberate, motivated camera motion for every shot—including subtle tracking during dialogue—rather than applying superficial visual effects.
Dolly zoom is graded NONE, and what we cost out instead
Dolly zoom is graded NONE for us. The prompt-only version has not been run. The route we planned instead is a grey Blender blockout passed in as a camera-path reference, or the Magnific 3D Motion dolly-zoom preset at +400 credits (2,600 versus 2,200 at 5 s, 720p).
Complex optical phenomena like the Vertigo effect (dolly zoom) cannot be executed through text prompts alone. In our capability matrix, prompt-only dolly zooms are graded NONE and have not been deployed in production. Feasible production routes require passing a simplified grey Blender blockout as a geometric camera-path reference, or deploying the Magnific 3D Motion dolly-zoom preset at a premium of +400 credits (2,600 credits versus 2,200 credits for a 5 s 720p take).
Camera movement prompt syntax, previz geometry, and production cost metrics · measured 25 Oct 2026
Camera Move Category
Cinematic Execution Syntax
Previz / Technical Specification
Production Cost / Billing Rate
Linear Tracking Move
'dolly in' ending with 'No zoom'
Forward travel with physical parallax
$1.85 (4 s @ 720p, $0.4622/s rate)
Dialogue Scale Move
Slow dolly in/out, restrained handheld
Locked scale during spoken dialogue
$0.82 (480p diagnostic probe)
Previz Video Reference
Video-guided camera path input
Billed at 0.6x previz coefficient
$2.22 (4 s take with 4 s previz)
Optical Dolly Zoom
Grey Blender blockout or 3D preset
Background compressed by ~2.4x
+400 credits (2,600 vs 2,200 at 5 s)
Top-Down Camera Descent
~40° FOV from 7.5 to 10 m elevation
Avoids wall warping seen at 65° / 4.5 m
Close-ups use ~29 degrees (~29°)
Rack Focus Transition
Focus shift only after dialogue ends
Never shift focus on speaking mouth
Capped at 1-2 signature takes per ep
Measured on 2026-10-25; one main move allowed for 5-8 s clips; max two for 10-15 s.
Dolly direction and top-down field of view in Blender previz
In our Blender previz, dolly back with zoom in compressed the background by about 2.4x and made it loom. A dolly zoom that makes the room rush away is dolly IN with zoom OUT; we got the sign wrong once and rewrote the line.
Pre-visualization reveals mathematical camera relationships. In synthetic camera modeling, dollies backward paired with simultaneous focal zooms inward compressed background perspective by approximately 2.4x (2.4x), causing background architecture to loom menacingly. Conversely, forcing the room to rush away requires dollying IN while zooming OUT. Correcting directional signs in 3D previz prevented flawed prompt instructions.
Top-down previz: about 40 degrees of field of view from 7.5 to 10 m worked; about 65 degrees from 4.5 m bent the walls of the set, while close-up dialogue framing uses ~29 degrees.
Optical field of view governs architectural distortion. In top-down descending camera setups, configuring approximately 40 degrees of field of view from 7.5 to 10 m elevation captured clean vertical staging; widening the lens to approximately 65 degrees from 4.5 m caused severe geometric curvature across room walls. For close-up dialogue framing, our camera specifications standardize on ~29 degrees (~29).
Production economics, signature caps, and the probe roadmap
A silent 4-second 720p signature take costs $1.85 at our $0.4622 per second Seedance rate; the 480p probe of the same take is $0.82. Video used as a previz reference is billed at 0.6x, so the same take with a 4-second previz is $2.22.
Directing dynamic camera moves requires disciplined cost accounting. Under our ByteDance Seedance 2.5 API agreement at $0.4622 per second, rendering a silent 4-second 720p signature take costs $1.85 ($1.85). Conducting an initial 480p diagnostic probe costs $0.82. When conditioning the shot on a 4-second pre-visualization video reference, the reference input is billed at 0.6x, bringing the total cost of the take to $2.22.
Signature moves are capped at 1 or 2 per episode. An orbit or arc around a large face re-renders that face from many angles, so for those we keep the subject small, a silhouette, or facing away.
Complex rotational moves must be rationed. Studio guidelines cap signature moves at 1 or 2 per episode. Because sweeping orbital arcs around large facial close-ups force diffusion models to re-derive biometrics across extreme angles, prompts keep subjects small in the frame, cast in silhouette, or facing away from the camera.
Rack focus is allowed in a dialogue shot only after the line ends. The focus shift never starts on a speaking face.
Focal transitions require temporal segregation. When executing rack focus shots during dialogue scenes, the focus shift may only occur after the actor finishes their spoken line. Adjusting focal planes while lips are moving introduces severe facial distortion.
Six named moves are on our probe list (top-down descent, push through a gap, rack focus, whip pan, dolly zoom, time-freeze pull-back). None has been run as a standalone matrix; we cancelled that plan and will fill the rows from 480p episode drafts instead. Lip-sync under a dolly or handheld move is still unmeasured.
Our experimental roadmap tracks six prioritized camera moves: top-down descent, push through a gap, rack focus, whip pan, dolly zoom, and time-freeze pull-back. Rather than expending budget on isolated test grids, the studio captures these movements directly through 480p production episode probes, while lip-sync stability under moving camera takes remains an active measurement focus.
