Free Runway Prompt Generator: 24 Gen-4.5 Templates

TL;DR: A Runway prompt is a prompt plus a request shape. Gen-4.5 takes 2 to 10 seconds, a 1,000-character prompt, six ratios on image-to-video but only two on text-to-video, no negative prompt and no last frame. The 24 templates below each name the branch and settings they need.

One honesty note before anything else. Prompt Architects writes the prompt, not the video. Nothing on this page renders a clip, and nothing here is a Runway account. These are words you paste into Runway, whether you drive it from the web app or from the API.

Which Runway Model Does Your Prompt Actually Need?

Start here, because getting this wrong is how good prompts fail. Runway's POST /v1/image_to_video is not one endpoint with one body. Runway's own agent primer puts it plainly: on generation endpoints the request body is a discriminated union keyed on model, so the valid ratios, the allowed durations, the prompt length limit and which optional fields exist all change with the model you pick. There are twelve such shapes on image-to-video and eleven on text-to-video, as of the specification fetched on August 27, 2026.

That means a prompt written for Veo 3.1 inside Runway can be structurally illegal for Gen-4.5 in the same account, on the same endpoint, on the same day. It is also why "a Runway prompt" is a slightly misleading phrase. Runway is a router as much as a model. Its video model list runs to fourteen entries, only four of which are Runway's own: Gen-4.5, Gen-4 Turbo, Aleph 2.0 and Act-Two. The rest are Veo 3.1 and its fast variant, four Seedance branches, Hailuo 3, Grok Imagine 1.5, Gemini Omni Flash and HappyHorse 1.0. The words you write are largely portable between them. The settings are not.

If you drive any of this through the API rather than the web app, two mechanical details bite before your prompt ever gets read. Every request must carry an X-Runway-Version header, currently 2024-11-06, and requests without it fail outright. And generation is asynchronous: the response to your create call contains a task id, not a video. You poll the task until it reaches a terminal status, and only a succeeded task carries output.

Two rows in that table are the ones people trip over. Gen-4.5 is the only branch here that makes promptText mandatory in both modes, so an image-only generation is not a legal Gen-4.5 request. And Gen-4.5 is the only branch here that exposes seed while offering no negativePrompt, which is an unusual pairing: you get reproducibility but no exclusion channel.

Is Gen-4.5 Still Runway's Current Video Model?

Yes, and this is worth stating because the answer keeps being no for other vendors. Gen-4.5 landed on Runway Dev on February 10, 2026. On August 27, 2026 it is still on the current models page, it is still gaining capability rather than losing it, and Runway's July 30, 2026 sunset notice for the previous generation reads: "Upgrade from Gen-3 Alpha Turbo to Gen-4.5 (best quality) or Gen-4 Turbo (fastest)". The changelog contains no Gen-5 entry.

What did change recently is delivery, not the model. On August 20, 2026 Runway added professional output containers to Gen-4.5 through an outputFormat field: ProRes, PNG sequences, 10-bit SDR, and several true HDR profiles. That is a post-production feature, and it does not alter a single word of how you write the prompt.

Runway's help centre describes Gen-4.5 as offering text-to-video and image-to-video control "with support for additional inputs coming soon", and says the model "excels at understanding and executing complex, sequenced instructions". Both statements match the schema: there is no Gen-4.5 branch on /v1/video_to_video at all, and the sequencing claim is what makes timestamp prompting worth using.

What Belongs in a Gen-4.5 Prompt, and What Does Not?

Runway publishes different advice for each mode, and the difference is the whole craft of the model. For text-to-video, its guide says effective prompts contain visual descriptions and motion descriptions, and offers a starting structure: [Camera] shot of [a subject/object] [action] in [environment]. [Supporting component descriptions]. For image-to-video the advice inverts, because the image already carries composition, subject, lighting and style. Runway's line is blunt: "Effective image to video prompts focus almost exclusively on motion." Its suggested structure there is The camera [motion description] as the subject [action]. [Additional descriptions].

Three published claims are worth internalising because they contradict habits people bring from image models. Runway says "Structure and order are far less important than clearly conveying an idea and reducing ambiguity". It answers the question of whether prompt openings are weighted with "The order in which elements are introduced in a prompt do not matter." And on formatting it is unusually direct: "JSON prompts give the placebo effect of being more accurate", because "JSON formatting is ignored by generative models". If you keep a JSON prompt library for other video models, do not point it at Runway and expect the structure to do work.

The absence of a negative prompt is not an oversight, it is consistent with the rest of Runway's guidance: "Models respond better to descriptions of what you want to happen rather than what you don't want to happen." Our negative prompt support matrix has the cross-vendor version of this, and Gen-4.5 is filed there as not published.

There is one place inside Runway where a negative prompt for Gen-4.5 appears to exist and does not. The Model Router endpoint takes a model-agnostic input object that accepts a 6,000-character promptText and a 1,000-character negativePrompt, and the spec describes those fields as optional options the router maps onto whatever model it selects. Send that body, let the router land on Gen-4.5, and your exclusion list has nowhere to go. If a negative prompt is load-bearing for the shot, name a branch that has one rather than routing and hoping.

Runway also refuses to declare a winner between short and long prompts, which is more honest than most vendor guidance. It publishes a two-column table weighing a simple strategy (faster to iterate, more creative freedom, better for exploration) against a detailed one (less ambiguity, more consistency across generations, better for defined requirements), and states outright that detailed prompts are not always better. Its summary line is the one to keep: "The goal of detail is to remove ambiguity, not to control every pixel." Runway warns that multi-paragraph over-specification can produce unnatural results because the model struggles to honour every detail at once, which matters more here than elsewhere given the 1,000-character ceiling.

Two more rules that come straight from the schema. The prompt ceiling is 1,000 characters on both Gen-4.5 modes, which is roughly 150 to 170 English words, so every clause has to earn its place. And if your input image does not match the ratio you requested, Runway states that "the model will auto-crop your asset from the center to the aspect ratio parameter provided". That is a documented crop, not a rumour, and it will quietly reframe a product shot.

Text-to-Video: Six Templates for Commercial Work

All six require gen4.5 on /v1/text_to_video. Only two ratios are legal there: 1280:720 and 720:1280. Duration is a required integer from 2 to 10.

[01 · gen4.5 · text_to_video · 1280:720 · 6 s]
Medium shot of a ceramic pour-over cone on a pale oak counter, morning
light raking in from a window at frame left. A hand enters from the
right holding a gooseneck kettle and begins a slow spiral pour. Steam
lifts through the light. The camera pushes in gently as the bed of
grounds blooms and settles. Shallow focus, warm neutral palette, soft
window light, clean commercial photography.

[02 · gen4.5 · text_to_video · 720:1280 · 5 s · vertical social]
Vertical close-up of a running shoe on wet asphalt at dusk, city lights
smeared in the background. A runner's foot drives down into frame,
splashing a thin sheet of water outward. The camera trucks left to stay
level with the stride as two more steps land. High contrast, cool blue
and amber, rain-slick reflections, sports commercial look.

[03 · gen4.5 · text_to_video · 1280:720 · 8 s · establishing]
A symmetrical establishing shot of an empty open-plan office before
opening, monitors dark, chairs pushed in, low sun cutting through
vertical blinds and striping the floor. Dust drifts through the light.
The camera holds still for three seconds, then begins a very slow push
in down the centre aisle. Muted greys and warm gold, calm corporate
documentary tone.

[04 · gen4.5 · text_to_video · 1280:720 · 4 s · b-roll insert]
Extreme close-up of a fountain pen nib crossing heavy cream paper,
laying a wet black line that catches the light. The camera tracks right
alongside the nib at the same speed, holding the tip in the centre of
frame. Shallow focus, paper fibres visible, single warm desk lamp from
frame right, quiet editorial mood.

[05 · gen4.5 · text_to_video · 720:1280 · 7 s · hands and product]
Vertical shot of two hands folding a linen shirt on a dark walnut
surface, framed from just above the wrists. Diffused overhead light,
no visible faces. The hands smooth the fabric, fold the sleeves in,
turn the shirt, and fold it once more into a neat rectangle. The camera
performs a slow pedestal down, ending level with the folded shirt.

[06 · gen4.5 · text_to_video · 1280:720 · 10 s · long single take]
A wide shot of a shuttered market street at first light, awnings still
rolled, a single delivery bicycle leaning against a wall. A trader
enters from frame left pushing a crate trolley, crosses the frame, and
begins unlocking a shutter. The camera arcs slowly right around the
trolley as the shutter rises, revealing stacked fruit crates inside.
Continuous, seamless shot. Cool dawn light warming through, handheld
documentary feel, natural camera shake.

Image-to-Video: Seven Motion-Only Templates

These require gen4.5 on /v1/image_to_video, where six ratios are legal. Runway labels them 16:9 1280:720, 9:16 720:1280, 1:1 960:960, 4:3 1104:832, 3:4 832:1104 and 21:9 1584:672. Remember that promptText is mandatory on this branch, and that the prompt should describe motion rather than restate the picture. Our guide to handing a still image to a video model covers the wider pattern.

[07 · gen4.5 · image_to_video · 1280:720 · 5 s · product hero]
The camera executes a slow, controlled orbit to the right around the
product, keeping it centred in frame throughout. Highlights travel
across the surface as the angle changes. Nothing else in the scene
moves.

[08 · gen4.5 · image_to_video · 960:960 · 4 s · square feed post]
The camera pushes in slowly and steadily on the subject. A faint drift
of steam rises and dissipates. The background falls further out of
focus as the push continues.

[09 · gen4.5 · image_to_video · 832:1104 · 6 s · portrait subject]
The subject turns their head slowly towards the lens and holds the
look. The camera pulls back a short distance at the same time,
revealing more of the room behind them. Fabric shifts slightly with
the movement. Continuous, seamless shot.

[10 · gen4.5 · image_to_video · 1584:672 · 7 s · ultrawide landscape]
The camera trucks right at a constant, gentle speed, parallel to the
horizon. Cloud shadows travel across the ground from left to right.
Nothing in the foreground enters or leaves the frame.

[11 · gen4.5 · image_to_video · 1104:832 · 5 s · interior reveal]
The camera cranes down from its starting height and settles at table
level. As it descends, the near edge of the table enters the bottom of
frame and the ceiling leaves the top. Light through the window flickers
as if a branch is moving outside.

[12 · gen4.5 · image_to_video · 720:1280 · 6 s · transformation]
The material begins to change from the bottom of the frame upward,
turning gradually from matte to a wet, reflective surface. The camera
tilts up to follow the change as it climbs. The transformation completes
just before the shot ends.

[13 · gen4.5 · image_to_video · 1280:720 · 8 s · two-element interaction]
The camera holds a static wide frame. The figure on the left walks
slowly towards the object on the right, stops an arm's length away,
reaches out, and rests a hand on it. The camera does not move at any
point. Minimal subject motion only.

How Do You Control When Things Happen in a Shot?

With timestamps, and Runway documents the syntax rather than leaving it to folklore. Its text-to-video guide describes timestamp prompting as a way to control when specific actions occur, notes that it "may not be perfectly precise", and recommends pairing it with a natural-language prompt rather than replacing one. The image-to-video guide gives the compact form: [00:01] X occurs. [00:03] Y occurs. [00:04] Z occurs.

The practical constraint is arithmetic. Gen-4.5 tops out at 10 seconds, so every beat you schedule has to fit, and Runway's own advice is to consider how long an action would reasonably take. Three beats in four seconds reads as a montage; three beats in nine seconds reads as a shot. Our camera movement vocabulary is the reference for naming each move precisely, and the seven-part video prompt anatomy is the general structure these templates compress.

[14 · gen4.5 · text_to_video · 1280:720 · 9 s · sequenced beats]
A barista stands behind a counter with their back to the lens, working
an espresso machine in a narrow café. Warm tungsten light, brass
fittings, steam in the air.
[00:00 through 00:03] working the machine, back to camera
[00:03 through 00:05] turns towards the lens holding a cup
[00:05 through 00:07] sets the cup down on the counter edge
[00:07 through 00:09] camera pushes in to frame the cup

[15 · gen4.5 · image_to_video · 1280:720 · 8 s · sequenced reveal]
[00:00] The camera holds still on the closed box.
[00:02] Two hands enter from the bottom of frame and lift the lid away.
[00:05] The camera pushes in over the open box.
[00:07] The push slows and stops.

[16 · gen4.5 · text_to_video · 720:1280 · 6 s · crash zoom beat]
Vertical medium shot of a chef plating a dish on a dark counter, facing
away from the lens under a single overhead light.
[00:00 through 00:03] plating, facing away
[00:03 through 00:04] turns sharply towards the camera
[00:04 through 00:06] rapid crash zoom in to frame their face

[17 · gen4.5 · image_to_video · 1584:672 · 10 s · slow choreography]
[00:00] The camera begins a slow truck to the left.
[00:04] A figure enters at the right edge of frame, walking left.
[00:07] The camera matches their pace and holds them at frame centre.
[00:09] Both the camera and the figure slow to a stop.

Which Prompts Give You a Genuinely Still Shot?

The ones that say so three times, in Runway's own words. Its camera guide concedes that "Video models are designed to create immersive, high-fidelity motion, which can make achieving static shots challenging", most notably on establishing and wide landscape shots, and recommends describing the motion that should occur inside the scene while reinforcing stillness in natural language. The published reinforcement lines include "The camera is entirely motionless for the duration of the scene, with movement only occurring from the subject", "The locked-off camera remains perfectly still", and, for loops, "The camera must start and end on the exact same frame to create a perfect loop".

[18 · gen4.5 · text_to_video · 1280:720 · 6 s · locked-off]
A static wide shot of a laundrette at night, six machines along the back
wall, two ceiling tubes flickering slightly out of sync. One machine is
mid spin cycle. The spin slows and stops. The camera is entirely
motionless for the duration of the scene, with movement only occurring
from the machine and the lights.

[19 · gen4.5 · image_to_video · 960:960 · 4 s · seamless loop]
The camera must start and end on the exact same frame to create a
perfect loop. The locked-off camera remains perfectly still. Only the
surface texture shifts, drifting slowly and returning to its starting
position.

[20 · gen4.5 · image_to_video · 1104:832 · 5 s · near-still insert]
The locked-off camera remains perfectly still. Minimal subject motion
only: a slow breath, a single blink, nothing else. No camera movement
at any point. Continuous, seamless shot.

Runway's stated fallback, worth repeating because it is honest, is to stabilise the near-miss in a local editor rather than to keep regenerating.

When Should You Route Away From Gen-4.5?

Four situations, and each one has a different branch on the same endpoint. This is where a prompt library beats a prompt: the text changes very little, but the request shape changes completely. Our duration parameters by model table is the cross-vendor version of this decision.

[21 · veo3.1 · text_to_video · 1920:1080 · 8 s · needs a negative prompt]
promptText: A steady wide shot of an empty hotel corridor at night,
patterned carpet running to a lit exit sign at the far end. The camera
dollies forward at a constant speed. Doors pass on both sides. The shot
ends before reaching the exit.
negativePrompt: people, reflections, text, signage, timestamps

[22 · seedance2_5 · image_to_video · 1920:1080 · 20 s · long take, 1080p]
The camera begins on the near end of the workbench and trucks slowly
right along its full length, passing tools, offcuts and a half-finished
frame. Sawdust drifts through the light. Halfway along, a pair of hands
enters and picks up a chisel. The camera continues past them without
stopping and settles on the finished piece at the far end.

[23 · aleph2 · video_to_video · matches input · 2–30 s input · edit an existing clip]
Keep the camera move, the framing and the timing exactly as they are in
the source clip. Change the time of day to late evening: warm low sun
replaced by cool blue ambient light, practical lamps switching on in the
background, and longer, softer shadows throughout.

[24 · gen4_turbo · image_to_video · 1280:720 · 4 s · fast draft pass]
The camera pushes in slowly on the subject while the background falls
out of focus.

Template 21 exists because negativePrompt is a Veo 3.1 field inside Runway, capped at 1,000 characters, and it is not available on Gen-4.5. Template 22 exists because Gen-4.5 stops at 10 seconds while the Seedance 2.5 branch runs to 30. Template 23 exists because Gen-4.5 has no presence on /v1/video_to_video at all: that endpoint carries seven branches and Aleph 2.0 is the Runway-built one. Template 24 exists because gen4_turbo is the only branch here where promptText is optional, which makes it a cheap way to test whether an image animates well before you spend a Gen-4.5 generation on it.

Why Do Runway Generations Fail Instead of Erroring?

Because moderation is a task outcome, not an HTTP status. Runway's content moderation page states that a moderated request returns a task with a FAILED status, with detail in failure and failureCode, and that "Moderated generations have the same credit cost as successful generations." It also states that too many moderated requests will get an account suspended. For anyone billing a client, that is a budget line, not an edge case.

The other operational trap is the output itself. Runway's primer is explicit: "Output URLs are temporary. Download and store anything you intend to keep." The output docs put the window at 24 to 48 hours and advise against exposing those URLs in a product. A delivered link that dies over a weekend is a support ticket you wrote yourself.

Two prompt-side guardrails follow from Runway's published usage policy, last updated March 6, 2026. It prohibits "Attempts to create content in the style of a known, living artist" and "Use of an image, video, or audio of another person without their permission". Strip named living artists out of your style clauses and use documented terminology instead. Every template above does this deliberately.

Can You Use Runway Output in Paid Client Work?

Broadly yes, with a clause worth reading before you sign anything. Runway's Terms of Use, last updated May 11, 2026, say at section 4.4 that "The Company does not claim ownership of any of your Inputs or Outputs" and that "Subject to your compliance with the Agreement, the Company does not restrict your commercial use of your Outputs." Runway's help centre puts the same thing more plainly: "the content you create using Runway is yours to use without any non-commercial restrictions from us", with no attribution required.

The part people miss sits in the same clause. You also acknowledge that "Inputs and Outputs may be used by the Company to train and improve its AI models", and you grant a perpetual, worldwide, sublicensable licence for that purpose. Ownership and licensing are different questions, and for a client under an NDA the second one is the one that matters. Check the current wording at runway.com/terms-of-use rather than trusting this paragraph, including ours: terms change, and dated quotes go stale.

What Should You Reuse Across Every Runway Prompt?

The label, not just the text. Every template on this page carries its branch, endpoint, ratio and duration in a header line, because in Runway those four values are part of the prompt in every sense that matters. A prompt without them is a sentence you will have to re-derive settings for, every time.

Beyond that, four habits survive across all 24. Write motion for image-to-video and write both visuals and motion for text-to-video. Say what you want rather than what you do not, because there is no exclusion channel on this model. Reinforce anything that failed once, in natural language, rather than escalating to formatting tricks that Runway says it ignores. And record the seed when a generation works, because Gen-4.5 exposes one and most of its sibling branches do not.

Treat each of these as a versioned prompt template with the variable parts marked, and the settings stored alongside the words. That is the difference between a library and a scratchpad, and it is the same discipline that makes a prompt engineering practice repeatable rather than lucky.

The Short Version

Gen-4.5 is still Runway's current video model as of August 27, 2026, which makes it a rare version-numbered claim that survives checking. It runs 2 to 10 seconds, takes a 1,000-character prompt, requires that prompt in both modes, offers six ratios on image-to-video and two on text-to-video, exposes a seed, and gives you neither a negative prompt nor a last frame. Write to those constraints, label every prompt with them, and route the jobs that do not fit to the branch that does.