Black Friday

全プランの1年分が50%オフ

今すぐ見る
eroq

ブログ/prompts·2026年11月21日·8分·執筆:eroqチーム

Runway Gen-4.5 prompt guide — literal prompts, exact camera paths, silent clips

How to prompt Runway Gen-4.5 on eroq: one clause per layer, camera paths in words, start frames and seeds, with 10 example prompts from 2 to 10 seconds.


Runway Gen-4.5 is the engine for shots where the picture has to match a precise instruction: one camera path, one action, one light, rendered as asked. The rule that matters most is to write it literally and in order — subject, action, camera, light, one clause each — because Gen-4.5 follows what you wrote instead of filling the gaps with ideas of its own.

That makes it a different tool from the engines that improvise well. On Runway Gen-4.5 a vague prompt gives a vague clip; a precise one gives you very nearly the shot you had in your head. This guide covers the prompt shape it rewards, the camera and light words it follows, ten prompts to start from, and the settings on eroq.

What Gen-4.5 is good at, and what it leaves out

Good at:

  • Prompt adherence. When the prompt says the camera starts on the railing and ends on the skylight, that is the path you get.
  • Short beats. Its lengths start at 2 seconds, which suits inserts, reactions and the B-roll you cut between longer clips.
  • Animating a still. Set an image as the start frame and Gen-4.5 moves it without redrawing it.
  • Revisions. It takes a seed, so a take you like can be revised one phrase at a time instead of re-rolled.

What it leaves out, by design:

  • Sound. Gen-4.5 renders picture only. Plan the audio elsewhere: a voice-over from Voice, a song from Music, laid under the clip in the Editor.
  • Reference pictures. It reads none, so an @mentioned character rides as text. For an exact face, start from a frame of that character.
  • End frames. It takes a start frame, never an end frame. For a move between two fixed images, pick Kling 3.0, Hailuo, Seedance or Veo.
  • Formats. 16:9 and 9:16 only, at 720p.

Prompt anatomy

Four layers, in this order, one clause each:

[subject, concrete] · [one action, with its speed] · [camera: move, height, lens] · [light: source, direction, time of day]

A fifth, optional layer carries texture or mood — haze, grain, a colour cast — in a few words. Between 25 and 60 words is the useful range: shorter and the engine guesses, longer and the later clauses start competing with the first.

What each layer needs:

  • Subject: who or what, specific enough to picture. "A man" is a guess; "a man in his sixties in a grey cardigan" is a casting note.
  • Action: one verb, plus how fast. "Slowly lowers the letter" beats "reacts to the letter".
  • Camera: the move, the height and, if it matters, the lens. "Slow dolly-in at eye level, 50mm" is a camera layer; "cinematic camera" is not.
  • Light: where it comes from and what it is. "Single window light from the left, late afternoon" lights the scene; "beautiful lighting" lights nothing.

Camera paths in words

Gen-4.5 is at its best when the camera has a start and an end. Write the path as a sentence: the camera starts on X and ends on Y. Words it follows well:

  • Moves: slow dolly-in, pull-back, lateral slider from left to right, low tracking shot alongside, crane up, crane down, slow pan, static locked-off, handheld at shoulder height, top-down overhead.
  • Speed: slow, steady, gradual, sudden. One speed word per move.
  • Height and angle: eye level, hip height, low angle, high angle from a ledge, overhead.
  • Lens: 24mm wide, 35mm, 50mm, 85mm close-up, macro.

The Syntax menu in the composer inserts the studio's 24 camera moves, and a move typed in the prompt is highlighted and applied. Lighting and Palette in Settings are folded into the prompt on our side, so the light layer can stay short when the rack carries it. The full list of moves, with when to use each, is in the camera moves guide.

Ten example prompts

Each one is written for Gen-4.5's literal reading: one subject, one action, one move, one light.

1. Reaction insert. A 3-second cut-in between two longer clips.

A chef in a white jacket looks up from the pass, eyebrows rising slowly. Static camera at eye level, 85mm close-up. Warm tungsten kitchen light from the left, steam drifting through the frame.

Settings: 3 s, 16:9, 720p.

2. A path with a start and an end. The camera layer does all the work.

An empty concrete stairwell painted pale green. The camera starts at the ground-floor railing and cranes straight up through the centre of the stairwell, past every landing, ending on a skylight full of grey cloud. Cold, even overcast light.

Settings: 8 s, 9:16, 720p.

3. A silent product turn. No sound needed, so nothing is lost.

An unbranded matte black wristwatch on a slab of dark slate. A slow lateral slider moves from left to right as a strip of hard white light crosses the dial and catches the steel hands. Macro lens, black background, crisp reflections.

Settings: 5 s, 16:9, 720p.

4. Tracking at the subject's pace.

A woman in a red running jacket jogs along a seaside promenade at sunrise. Low tracking shot alongside her at hip height, matching her pace. Low sun behind her, long shadows, light haze over the water.

Settings: 6 s, 16:9, 720p.

5. Animate a still. With a start frame, describe only what moves.

The curtains begin to sway in a breeze and the candle flame leans toward the window. The camera holds still.

Settings: 4 s, the format of your still, start frame set.

6. Weather as the action.

Heavy rain hammers a deserted bus shelter at night. Static wide shot from across the street. Orange sodium streetlight, wet asphalt reflecting it, a single car passing and throwing up spray.

Settings: 6 s, 16:9, 720p.

7. Hands at work, from above.

Top-down overhead shot of hands kneading bread dough on a floured wooden table, folding and pressing it in a steady rhythm. Soft window light from the top of the frame, flour dust in the air.

Settings: 5 s, 9:16, 720p.

8. A reveal by pull-back. One move, and the reveal is the point.

Close on a climber's chalked hand gripping a granite edge. A slow pull-back reveals her hanging above a vast valley at golden hour, wind tugging at her sleeves.

Settings: 8 s, 16:9, 720p.

9. A performance without a word.

An old man sits alone at a chess board in a park, reaches for a knight, hesitates, and draws his hand back. Static medium shot at eye level, 50mm lens, dappled afternoon light through plane trees.

Settings: 5 s, 16:9, 720p.

10. B-roll for a voice-over.

Morning commuters cross a wide intersection from both sides, seen from a high ledge. Slow pan from left to right. Pale spring light, long shadows, city haze.

Settings: 8 s, 16:9, 720p, with the voice-over laid under it in the Editor.

Revise with the seed, not with luck

Once a take is close, open Settings and copy its Seed. Keep it, change one thing — the lens, the speed of the move, the light — and generate again: the next take changes on that one axis. Without a seed every take is a fresh draw, and you can't tell whether your edit helped. Leave the field empty when you want variety, and use Takes to render up to four at once; identical takes with the same seed look the same.

Mistakes to avoid

  • Asking for sound or dialogue. Gen-4.5 renders none. Lines in quotes come back with mouths that may move and no voice. Record the line in Voice and cut it in the Editor, or pick an engine with a soundtrack.
  • Vague camera words. "Dynamic shot", "cinematic camera" and "epic angle" give the engine nothing to follow. Name the move, the height and the speed.
  • Several actions in one clip. "She opens the door, crosses the room and sits down" is three beats; ten seconds will rush all three. Render one and put the next in the next clip.
  • Describing the subject again on a start frame. The frame already carries the face, the wardrobe and the set. Re-describing it invites the engine to redraw them.
  • Saying what you don't want. "No people in the background" puts people in the prompt. Describe what is in frame instead: "an empty street".
  • Expecting a cast to hold. No reference pictures ride on this engine. For a recurring face, render a still of the character in Image, then use it as the start frame.

Settings on eroq

  • Where: Video › Create, then Runway Gen-4.5 in the engine list.
  • Duration: 2 to 10 seconds, in one-second steps.
  • Resolution and Format: 720p, in 16:9 or 9:16.
  • Inputs: text, a start frame and a seed. No end frame, no reference pictures, no soundtrack.
  • Price: 51 credits / second — 102 credits for a 2-second insert, 255 credits for 5 seconds, 510 credits for 10. The Generate button shows the total before you commit, and a failed render refunds itself.
  • Plan: Creator ($49/month) and up.
  • Privacy: Runway Gen-4.5 is moderated upstream by its provider, which keeps its own request logs. A prompt it refuses returns content_blocked and costs nothing.

For a whole story, storyboard it in Cinema and give Gen-4.5 the precise shots — the inserts, the reveals — while engines with sound carry the dialogue scenes.

FAQ

Is Runway Gen-4.5 free to try on eroq?

eroq is free to start: 40 free credits when you sign up, no card. Runway Gen-4.5 itself opens on the Creator plan ($49/month) at 51 credits / second; the free credits run the engines that need no plan, such as Seedance 2.0 Mini, Veo 3.1 Lite and Motion One.

Does Runway Gen-4.5 generate sound?

No. It renders picture only. Add a voice-over from Voice or a song from Music, and lay it under the clip in the Editor.

How do I keep the same character across Gen-4.5 clips?

Render a still of the character in Image — from their Character, with their face — and set it as the start frame of each clip. Gen-4.5 reads no reference pictures, so the start frame is what carries the face.

Why does Gen-4.5 ignore part of my prompt?

Usually because the prompt holds more than one action or more than one camera move, and the later ones lose. Cut it to one subject, one action, one move and one light, then add detail back one clause at a time with the seed held.

Pick Runway Gen-4.5 in the engine list and start with a 3-second insert — /studio/video.

タグprompt-guiderunwaytext-to-videoimage-to-video

Cinemaで編集する

ここまでの内容はすべてスタジオで再現できます。同じエンジン、同じ設定、レンダリングはあなたのクリックで始まります。

40クレジット無料 · カード不要 · 全アカウントでプライベートモード

この記事のプロンプト

A chef in a white jacket looks up from the pass, eyebrows rising slowly. Static camera at eye level, 85mm close-up. Warm tungsten kitchen light from the left, steam drifting through the frame.

プライバシー

プライベート、そして制限なし

  • →eroq独自のエンジンにフィルターはなく、明文化されたポリシーがひとつあるだけ。
  • →何も保持せず、あなたの作品を学習に使うこともありません。
  • →プライベートモードはファイルを一切残しません。