Runway Gen-4.5 Prompts for Characters That Don't Drift (2026)

Most Runway Gen-4.5 prompts fail before they reach the model — not from being too short or too long, but from describing a feeling instead of a shot.

"Cinematic shot of a woman walking through a desert at sunset, beautiful lighting, dramatic mood" reads like a prompt. It isn't one. Every word asks the model to guess. Which lens? Which direction does she walk? What does "dramatic" mean in light angle and shadow length? Gen-4.5 doesn't reliably translate mood words — cinematic, dramatic, beautiful — into specific decisions. It reads camera direction, subject action, and lighting geometry. Everything else gets averaged into something generic.

That's the entire gap between flat, characterless Gen-4.5 output and shots that hold together across a sequence.

The fix: write shot directions, not descriptions

Replace every mood adjective with a named camera move, a named light source, and a beat-by-beat subject action.

Weak: Cinematic shot of a woman walking through a desert at sunset, beautiful lighting, dramatic mood.

Strong: Low-angle tracking shot, camera moving backward at walking pace. Woman strides forward, arms relaxed, gaze fixed ahead. Sun low on the horizon behind her, rim-lighting her silhouette. Long shadows stretch toward camera.

The second version gives Gen-4.5 a lens height, a movement, a gait, a light source, and a shadow behavior. Nothing left to average.

Keeping a character consistent across shots

Gen-4.5 carries over the @ reference system from Gen-4. Upload a reference image, tag it with @, and the model holds that character's appearance across separate generations — same face, same outfit, same proportions, shot after shot.

This matters more than any single prompt trick, because character drift is the fastest way a sequence falls apart. A face that shifts slightly between clips reads as "AI-generated" even when each individual frame looks clean. Lock the reference first. Write the shot direction second.

For dialogue with a synchronized facial performance layered on top, Act-Two is the companion model — it takes your reference and drives a real performance from a source video or set of instructions, not just holding an appearance still.

Working inside the 10-second limit

Gen-4.5 generates in 5- and 10-second clips. Treat each generation as one scene, not a mini-movie. Prompts trying to pack in a scene change, a subject transformation, and a style shift within one clip force the model to reconcile contradictory instructions — and it shows.

The workflow that actually holds up in production:

1. Draft fast on Gen-4 Turbo to test motion and blocking cheaply
2. Final pass on Gen-4.5 once the shot direction is locked, same seed
3. Edit with Aleph for small adjustments — relighting, background swaps, reframing — without regenerating from scratch
4. Layer performance with Act-Two if the shot needs dialogue or facial nuance

Skipping the Turbo draft and going straight for a finished shot on the first pass is where most people burn through credits chasing a result iteration would have reached faster.

Mistakes that waste generations

• Negative phrasing. "The camera doesn't move" confuses the model more than it helps. State what the camera does instead.
• Stacking aggressive camera moves at high values in one prompt. Start smaller than you think you need.
• Skipping the reference upload when consistency matters, then trying to fix drift in the prompt text alone.
• Treating it as one-and-done. The real power is the workflow — generate, edit, perform, finish — not a single perfect prompt.

Build the full shot in under two minutes

Getting this right by hand means holding camera terminology, lighting vocabulary, and reference-tagging syntax in your head every time you sit down to prompt. The AI Runway Gen-4.5 Pro Prompt Builder turns that into 32 dropdown fields — camera movement, lens height, lighting geometry, subject action, and Aleph/Act-Two-specific outputs — and generates five ready-to-paste prompts: a main Gen-4.5 shot prompt, an Aleph edit brief, and an Act-Two performance brief, all consistent with the same reference character.

Pairs well with

• AI Veo 3 Video Prompt Builder — for shots that need native audio generated alongside the picture
• AI Cinematic Video Prompt Builder — for shot-list and color-grade planning across a full sequence
• AI Product Video Prompt Builder — for applying the same shot-direction discipline to commerce listing videos
• AI Hailuo H3 Prompt Builder — for MiniMax Hailuo H3 shots with camera movement and motion strength control

Browse the full AI Video Prompt Builders collection.