Two levers: structure vs style
Describe the visual setup (subject, environment, lighting, composition, and style) separately from the motion (subject action, camera movement, environmental movement, direction, and speed). This makes it easier to identify which part of a failed generation needs to change.
Prompt density
- Start with the essential scene and motion. Add palette, grain, focal distance, or atmospheric detail only when it helps the intended result.
- For stylized overlays on UI screen recordings, obsess over text legibility invariants—explicitly forbid warping monospace columns.
- Honor duration discipline: latent tools often bill coarse time increments—sketch arcs in beats that survive trimming.
Reference frames
When injecting an image-conditioned first frame, align composition with genuine frame zero whenever continuity matters—misaligned starters produce jump-cuts corrected only by expensive re-generation or manual edits.
Dailies workflow
Batch generate short loops before assembling long arcs. Apply single-change critiques (“raise contrast of rim light”, “suppress particle density in foreground occlusion”) analogous to iterative GPT Image edit cadence—drift compounds across chained latent steps.
Source and further reading
Runway’s current text-to-video prompting guide recommends direct language, separating visual and motion components, and adding detail only as needed.