๐ฅ 30 Runway Gen-4.5 Prompts: References & Camera Templates
Motion-first prompt templates for Runway Gen-4.5's @ references, grouped by use case: character consistency, camera moves, sequential action, restyle recipes.
Jordan Reyes ยท AI Video Producer
ยท 5 min read
โก TL;DR โ quick answers
- What's different about prompting Gen-4.5 versus Gen-4?
- The reference system carried over โ the @ syntax for tagging up to three reference images is unchanged. What's sharper in 4.5 is motion adherence: the model reads sequential instructions ('X happens, then Y, then Z') more reliably, so prompts that describe an order of events hold up better than they did on Gen-4.
- Should I describe what's already in my reference image?
- No โ that's the single most common mistake we see. Runway's own guidance is explicit: the image already defines composition, lighting, and style. Re-describing the subject wastes prompt space that should go to describing motion. Say what changes, not what's already visible.
- How many reference images can I use at once?
- Up to three per generation with the @ syntax. Tag each one, then describe what it contributes: a character reference, a setting reference, a style reference, rather than letting the model guess which reference does what.

Runway's own prompting guidance says it plainly: describe movement, not what's already visible in the frame. Most bad Gen-4.5 prompts break that one rule. They re-describe the reference image in loving detail and leave the model to guess what should actually change. These 30 templates, grouped by the situations we hit most, are built around that rule instead of around it.
By the numbers
- Gen-4.5's References system accepts up to 3 reference images per generation via the
@tag syntax carried over from Gen-4 - Runway's own guidance: prompts that lead with motion description over subject re-description consistently produce cleaner adherence, per the official Gen-4.5 prompting docs
- Sequential language ("X occurs, then Y, then Z") is the documented method for temporal control โ not hard timestamps, which the model has to fight against rather than follow
The one rule that fixes most bad prompts
Tag your reference, name what it contributes, then spend the rest of the prompt on what moves. That's it. @character isn't a caption for the image. It's a pointer, and the prompt's job starts after the pointer, not before it.
Character consistency (6)
@characterwalks into frame from the left, turns to face camera, slow push-in, natural daylight@characterreaches for a cup on the table, picks it up, takes a slow sip, static medium shot@characterturns their head toward an off-screen sound, brief pause, handheld micro-shake@characterand@locationโ subject enters the room, crosses to the window, looks out, soft afternoon light@charactersits down slowly into frame, settles, exhales visibly, locked-off wide shot@characterstands, adjusts their jacket, then walks toward camera and past it, tracking follow
Camera-move templates (6)
- Slow push-in on
@character's face as their expression shifts from neutral to a small smile - Camera orbits right around
@characterstanding still, revealing the space behind them - Handheld follow behind
@characterwalking down a corridor, slight sway, practicals flickering past - Static wide shot,
@charactercrosses frame left to right at a relaxed pace, then exits right - Slow pull-back reveal from a close detail on
@character's hands to a full wide shot of the room - Whip-adjacent tracked pan (not fast-cut) following
@characteras they turn to walk away
Sequential / multi-beat action (6)
- First
@characterpicks up the phone, then pauses, then sets it back down without answering @characteropens the door, steps through, then turns to close it behind them- The rain starts lightly, then picks up, then
@characteropens an umbrella and looks up at the sky @characterreaches for the glass, hesitates, then pulls their hand back and looks away- The candle flame flickers once, then steadies, then
@characterleans in and blows it out @characterstands from the chair, walks to the window, then rests a hand on the frame and stops
Restyle / style-reference recipes (6)
@characterand@styleโ same pose and action, rendered in the reference's color grade and film grain@locationreimagined in@style's lighting language โ same geometry, different mood, slow static hold@characterwalking, matched to@style's handheld camera texture and lens distortion@producton a turntable, lit and graded to match@style's reference palette, slow orbit@characterspeaking to camera, restyled with@style's grain and contrast, static medium shot@locationat golden hour, restyled toward@style's cooler, more desaturated look, slow drift
Product / object motion (6)
@productrotates slowly on a dark surface, single key light sweeping across the label, static camera@productsits still as a hand enters frame, picks it up, and turns it toward camera@producton a reflective surface, camera does a slow low-angle push-in, studio lighting@productunboxing: lid lifts first, then the product is revealed, then a hand lifts it out@productfloats into frame from below, settles, then a soft light sweep passes across it@productand@locationโ the object sits on a table as the camera pulls back to reveal the room around it
Why the sequential ones work better than they should
Gen-4.5 doesn't need a timestamp to understand order. It needs the order stated in plain language. "First the door opens, then she steps through, then the camera holds" gives the model three beats without demanding it hit a rigid clock. We tried the timestamp version of the same prompts early on ("0-2s: door opens, 2-4s: character enters") and got worse adherence, not better โ the model spent effort trying to satisfy a timing constraint instead of the action itself. Drop the clock, keep the sequence.
What still breaks these
Stacking more than three beats into one prompt is the most common failure we see. Gen-4.5 handles two or three sequential actions cleanly and starts dropping or blending steps past that. If a shot needs five beats, that's two generations, not one overloaded prompt. Conflicting reference and prompt information is the second failure: if @style shows warm tungsten light and the prompt says "cool blue moonlight," the model splits the difference into something that satisfies neither.
How we picked
These templates are built directly from Runway's own Gen-4.5 prompting documentation and the official Creating with Gen-4.5 guide โ the @ reference syntax, the motion-over-description principle, and the sequential-language method for temporal control are all documented Runway guidance, not reverse-engineered guesses. Category groupings and specific prompt wording are ours, shaped by the References workflow we use in production. For the underlying model's specs and pricing, see our full Runway Gen-4.5 guide; for when Runway's toolkit loses to Kling's motion control, our comparison covers where each one wins.
Three reference images and a sentence about what changes beats a paragraph re-describing what's already on screen. Every prompt above is built on that trade.
Frequently asked questions
โธWhat's different about prompting Gen-4.5 versus Gen-4?
The reference system carried over โ the @ syntax for tagging up to three reference images is unchanged. What's sharper in 4.5 is motion adherence: the model reads sequential instructions ('X happens, then Y, then Z') more reliably, so prompts that describe an order of events hold up better than they did on Gen-4.
โธShould I describe what's already in my reference image?
No โ that's the single most common mistake we see. Runway's own guidance is explicit: the image already defines composition, lighting, and style. Re-describing the subject wastes prompt space that should go to describing motion. Say what changes, not what's already visible.
โธHow many reference images can I use at once?
Up to three per generation with the @ syntax. Tag each one, then describe what it contributes: a character reference, a setting reference, a style reference, rather than letting the model guess which reference does what.
โธDo I need timestamps in a Gen-4.5 prompt?
No, and adding hard timestamps tends to fight the model rather than help it. Sequential language, 'first the door opens, then she steps through, then the camera holds,' gives Gen-4.5 the same ordering information without forcing a rigid timing the model has to fight to hit.
The 5 best AI video finds, every week
New models, tested prompts, and what actually worked in our production โ one short email a week. No spam, unsubscribe anytime.

Written by Jordan Reyes
AI Video Producer
Runs multiple faceless YouTube channels and tests every major AI video model against the same prompts before recommending one. Tracks render time and credit cost like other people track calories.
Explore these topics
Every guide, comparison and prompt library we have on each.





