Home›Guides› Duration & Aspect: The Invisible Instruction

Duration & Aspect: The Invisible Instruction

7 min read

Clip length and frame shape are self-fulfilling instructions. Practical notes from our corpus benchmarks.

Frame dimensions and temporal length are not post-production containers — they directly shape how diffusion models organize visual weight and movement momentum. A model does not render a square and crop it to your delivery format. It initialises a latent tensor grid matching the ratio you asked for, and the attention budget gets distributed across that grid from frame one.

01. Stating Duration Explicitly

Models do not count seconds like human editors, but temporal conditioning weights respond heavily to numeric cues like “5-second clip” or “continuous 10s take”. Placing duration near the prompt boundaries reinforces temporal coherence.

02. Aspect Ratio as Compositional Steering

  • 9:16 (Vertical): Naturally prioritizes single subjects, vertical architecture, and intimate eye-level portraits.
  • 16:9 (Landscape): Prioritizes horizontal kinetic movement, group interactions, and expansive horizon lines.
  • 2.39:1 (Cinemascope): Forces wide panoramic staging with strong edge-to-edge environmental context.

Never rely on cropping an output in an external editor when you can condition the model natively to compose within your target aspect ratio from frame zero.

03. Why Native Ratio Beats Post-Cropping

Rendering 16:9 and cropping the centre for a vertical feed reliably ruins the result. The model staged its composition for a wide attention grid — subjects sat off-centre, environmental storytelling lived in the wings you just cut away, and rescaling a horizontal render to vertical throws away roughly 44% of the pixels you paid to generate. Every time you crop, you are removing content the model placed deliberately.

Generate natively at 9:16 when 9:16 is the delivery target. The vertical grid compresses attention horizontally, which actually helps single-character staging: the model has fewer places to scatter a face.

04. Duration Degradation Curve

Fidelity does not stay flat as clips lengthen. The useful pattern across the current frontier models:

  • 3–6s — highest fidelity. Sharp micro-texture, clean physics, no drift. Best window for anything you will cut into a larger piece.
  • 8–12s — minor background drift accumulates. Static environmental elements (foliage, distant crowds, reflections) begin to soften or rearrange between frames.
  • 15s+ — multi-action prompts start hallucinating. Extra limbs, warped architecture, and duplicated characters are common unless the camera motion stays extremely simple.

The practical consequence: generate in short shots and cut them together. Five 6-second clips edited into a 30-second sequence will outperform one 30-second generation almost every time, because each 6s render is inside the high-fidelity window.

05. Why the Camera Must Be Simple at Length

The degradation compounds with camera complexity. A slow push-in holds a stable subject and can survive 12 seconds. A 15-second handheld orbit around a walking character asks the model to maintain 3D consistency and independent limb motion simultaneously, and it will fail.

Match the camera to the duration, not the other way around:

DurationSafe camera movesAvoid
3–6sAnything, including fast handheld—
8–12sSlow pushes, pans, static tripodFast orbits, whip pans
15s+Static, or one very simple driftAnything multi-axis

06. Duration Near the Prompt Edges

Temporal conditioning decays with distance from the conditioning boundary. If duration is the last thing in a long prompt, it carries less weight than if it sits early. Put the output parameters — 8s, 16:9 — at either the very start or the very end, not buried in the middle of your descriptive clauses.

Worked Example

Vertical portrait format: slow smooth tilt up from a pair of worn leather hiking
boots resting on a stone mountain ledge to reveal a young mountaineer gazing out
over sunset clouds, golden hour backlighting, gentle wind in hair, 8s, 9:16.

Note what the ratio does for free here. A 9:16 grid is ideal for a tilt — vertical camera movement is the one axis with room to move in a portrait canvas. The same prompt at 16:9 would have forced you to reframe the reveal as a pull-back instead, or accept a subject lost in the horizontal negative space.


Reading list: more guides, the techniques manual, or dive straight into the library.

Fresh clips

Prompts to practice on

Open library →