How to Actually Get Usable Shots Out of Runway Gen-4

1. Let the image do the looking, and the prompt do the moving

This is the single biggest mistake I see, and it wastes more credits than the next five combined.

People type a full scene description into the prompt box, “a woman in a red dress standing in a sunlit café, warm afternoon light, cinematic, moody”, while the reference image already shows exactly that woman in exactly that café. That prompt is dead weight. The input image is required on both Gen-4 models. The input image establishes the visual starting point of the entire generative process and acts as the first frame of your output video. After uploading, selecting, or creating the input image, begin drafting your text prompt. Since the image conveys key visual information about subjects, composition, colors, lighting, and style, your text prompt should be almost entirely focused on describing the desired motion.

Translation: don’t restate the picture. The picture is the picture. Your job in the prompt box is to tell the model what changes over time. Camera moves. Subject actions. Light shifts. Wind. Steam. A hand rising. A head turning. If your prompt could be pasted next to a photograph and describe the still image accurately, you’ve written the wrong prompt.

Rewrite this: “A woman in a red dress in a sunlit café.” As this: “The woman lifts her coffee cup and turns her head slowly toward the window. The camera pushes in on a slow dolly.”

Same image. Vastly different output.

2. Write in short, plain sentences, not keyword soup

If you’re coming from Midjourney, you’ve got a bad habit to unlearn immediately. Runway doesn’t want a comma-separated list of vibes.

While external LLMs thrive on natural conversation, Runway’s models are designed to thrive on visual detail. Conversational elements like greetings or explanations waste valuable prompt space. Similarly, command-based prompts that request changes often lack the descriptions needed to convey how an element should behave in the output. For example, rather than directly asking to add or remove elements, instead describe how the elements should appear or disappear from the scene.

In practice: “cinematic, moody, 8k, dramatic lighting, ultra detailed” tells Gen-4 nothing useful. “The camera slowly pushes in as the streetlight flickers once and the rain intensifies” tells it everything.

And keep it short. The Gen-4 model thrives on prompt simplicity. Rather than starting with an overly complex prompt, we recommend beginning your session with a simple prompt, and iterating by adding more detaiil as needed. Start minimal, see what the model gives you, then add. The reflex to front-load every possible detail into the first prompt is exactly backwards.

3. One scene, one primary motion, you don’t have 30 seconds, you have five

Every Gen-4 clip is a single shot. Not a scene, not a sequence, not a mini-movie. A shot.

Gen-4 generates videos in 5 and 10 second clips, so it can be helpful to consider each generation as a single scene. Attempting to dictate each second of the video with multiple scene changes, subject actions, or style shifts may provide unintended results as the model attempts to reconcile too many disparate elements or contradictory instructions. In most cases, a simple description of the desired motion for a single scene will work well and allow the model to shine.

If your prompt has the word “then” in it, you’ve probably already lost. “The character walks in, then sits down, then picks up the phone” is three shots, not one. Pick one. Generate it. Move on. If you want the whole sequence, storyboard it as three separate generations and cut them together.

The other side of this rule: one primary motion, one secondary motion, and stop there. Camera does one thing. Subject does one thing. Environment does one thing. That’s it. Cram in more and the model starts hedging, and hedging looks like warping.

4. Name the camera move like a DP, not like a hype reel

Gen-4 has clear opinions about camera language, and it respects real filmic vocabulary. Use it.

Words that work: push in, pull out, dolly, truck left, truck right, orbit, pan, tilt, tracking shot, static, handheld, low angle, overhead. Tracking: Follows the subject’s movement horizontally or vertically. Panning: Rotates the camera on its horizontal axis. Tilting: Moves the camera up or down. Dolly: Moves the entire camera forward or backward. Handheld: Creates a slightly shaky, naturalistic feel.

Words that don’t work as well: “cinematic camera movement,” “dynamic shot,” “epic angle.” Those are marketing words. Gen-4 doesn’t know what to do with them, so it defaults to something safe and slightly floaty.

Also: avoid combining conflicting styles like “handheld” and “cinematic” in one sentence. Pick a camera personality. “Locked-off tripod” and “handheld” are different DPs. Don’t hire both for one shot.

5. Prototype in Turbo, finish in Gen-4

This is the credit-saving rule, and almost nobody follows it.

The Turbo model uses 5 credits per second of video · The Gen-4 model uses 12 credits per second of video · We recommend testing generations in Turbo, then switching to Gen-4 as needed. Turbo is more than half the price. Use it for everything that isn’t final.

Choosing between Gen-4 and Turbo: Because of the reduced credit cost and faster generation times, we recommend first exploring generative possibilities in Turbo. If you’re finding that your results in Turbo aren’t quite meeting your expectations, switching to Gen-4 may help you achieve your desired outcome.

My actual workflow: I dial in the reference image, the prompt, and the motion in Turbo, usually four to eight tries at five seconds each. Once the shot is 90% there, I lock the seed and run the final in Gen-4 at ten seconds. That’s the pipeline. Anything else is throwing money at the model.

6. Say what you want, not what you don’t want

Gen-4 doesn’t do negative prompts the way image models do. Trying to sneak them in via natural language often backfires spectacularly.

For the image model specifically: Negative prompts, or prompts that describe what shouldn’t appear in the image, are not supported in Gen-4 Images. Including a negative prompt may result in the opposite happening. Yes, really, write “no crowd in the background” and you may well get a crowd.

The fix is the same one that works everywhere else in Gen-4: describe the positive. Instead of “no other people in the frame,” write “a single figure alone at the far end of the platform.” Instead of “no camera shake,” write “the camera holds completely locked, mounted on a tripod.” Instead of “not cartoonish,” feed a photographic reference image and say “photorealistic, natural skin texture.” Positive phrasing gives the model something to build; negative phrasing gives it something to fight.

7. Use References for identity, and label them when you use more than one

Gen-4 References is the feature that separates hobbyists from people producing usable brand and narrative work, and most people still aren’t using it right.

The idea is simple: an identity, a face, a product, a location, that has to stay consistent across multiple shots lives in a reference image, not in your prompt. The model supports up to three reference images and accommodates various resolutions, with a maximum of 720×720 pixels for 1:1 and 1280×720 pixels for 16:9 formats. The reference is your anchor. The prompt is your director.

If you’re using more than one reference in a single generation, label them. Using consistent image labels such as “image_1”, “image_2”, and “image_3” in your prompts allows Runway Gen-4 to clearly understand which inputs should influence the output, and how. This approach is essential when combining elements across different references or when compositional control is important.

A prompt like “combine the character from image_1 with the environment from image_2, and place the object from image_3 on the desk” gives the model an unambiguous map. Dumping three reference images and writing “a woman in a room with a lamp” is asking the model to guess which reference is which. It’ll guess wrong, and the credits are still gone.

One more reference discipline: don’t fight your own reference. Over-Describing Reference Features: Avoid restating what’s already in the reference. Conflicting Instructions: Don’t change the reference subject in incompatible ways. If the reference shows a woman with brown hair, don’t write “same character but with blonde hair.” That isn’t a reference, that’s a contradiction. Generate a new reference in the image model first, then bring it in clean.

A bonus, because it’s the habit that ties it all together

Lock your seed the second a shot works.

You’ll get a take you love, the framing, the light, the pacing, all of it, and your instinct will be to keep pushing. Don’t. Save the seed. Then, and only then, change one variable at a time: bump the motion intensity, swap the aspect ratio, extend from 5 to 10 seconds. If you rewrite the whole prompt from scratch, you’ll never learn which lever actually moved the result, and you’ll never reproduce a look across a sequence of shots.

The mindset shift that fixes everything: stop thinking of Gen-4 as a text-to-video model with an image attached. Think of it as a cinematographer you’re briefing on set. The reference image is your location, your cast, and your production design, already scouted, already lit. The prompt is the one thing left: what happens in this shot, and how the camera behaves while it happens. Brief it that way and Gen-4 stops feeling like a slot machine. It starts feeling like a crew.