Runway Gen 3 has become a reference point for creators who want AI video that actually looks like film. Earlier tools produced recognizable but waxy images and motion that felt automatic. Gen 3, by contrast, rewards deliberate prompting. The difference between a generic AI clip and something that reads as cinematic is rarely the model itself; it is how carefully you describe what you want the camera, the light, and the subject to do. This tutorial is a practical guide to writing photorealistic prompts that hold up, built around the exact craft techniques that separate one-shot luck from a repeatable skill.
We will cover prompt structure, the vocabulary of lighting and camera, negative prompts, and the workflows needed to keep a scene and its characters consistent across multiple shots. The advice is organized so you can apply it immediately and keep it in mind as the platform evolves.
Launch cinematic AI video
Cinematic, in the AI video sense, does not mean gimmick. It means that every element of the frame reads as intentional: the light has a source and a quality, the camera has a language, and the subject occupies space with physical believability. Gen 3 can deliver all of this, but only if the prompt gives it a coherent world to render.
The most common beginner mistake is treating the prompt as a description of what is on screen. A prompt like "a robot walking through a city" tells the model nothing about mood, light, or lens, so the output is a lottery. A cinematic prompt instead treats each field as a decision you are making as a cinematographer. Learning to think in those decisions is the whole game, and you can practice it before you ever open the app.
The fundamental principles of photorealistic prompting
Photorealism in AI video depends on anchoring the model in plausible physics and materiality. The model is better at realism when the prompt describes surfaces, light, and scale rather than vague adjectives. Words like "authentic," "realistic," or "high quality" do little; the model already assumes it should render well. Far more useful are concrete references to film grammars, camera hardware, and lighting setups, because those constraints guide the render into a coherent, believable space.
Think in terms of constraints that the model can honor. Describe a lens focal length, an aperture, a camera distance, and a light source. Describe materials by their texture, weight, and how they react to light. And describe motion not as a command but as a physical event, a hand lifting, fabric settling, dust drifting. Every concrete constraint you add shrinks the space of possible outputs toward the one you actually want.
Writing a structured prompt, block by block
Break your prompt into discrete blocks and assemble them in a stable order. Start with the shot and camera: a lens length like 24mm or 85mm, a distance, and an angle, such as "low angle, wide establishing shot, 28mm." Next define the subject and its action, naming what it is, its material, and precisely what motion it performs. Then choose the scene and environment, including scale, time of day, weather, and architecture. After that set the lighting: the source, its quality, and its direction, like "golden-hour side light, soft keys, long shadows." Give tone and color, such as "moody teal and amber palette, desaturated, high contrast." End with finish tokens that specify texture and grain, like "shallow depth of field, 35mm film grain, subtle vignette."
Keep this order consistent across a project. When the model responds well to one prompt, small edits carry forward predictably; when you change your subject, you only swap the subject block. This modular discipline means you build a repeatable process rather than starting from scratch for every shot, which is exactly how you scale beyond single lucky outputs.
The vocabulary of light and camera
Light is the fastest route to cinematic feeling. Name your sources and give them quality and mood: "hard afternoon sun," "soft diffused window light," "a single warm practical lamp against cold shadow." Decide on direction, so the model does not default to flat ambience, and let contrast express drama. A scene that is all flat light will never feel filmic no matter how good the model is.
Camera vocabulary is equally important. Focal length hugely changes the read of a frame: wide lenses exaggerate space and draw the viewer in; telephoto lenses compress and flatter. Aperture separates subject from background, and camera height creates psychology, low angles add power, eye level adds neutrality, high angles add vulnerability. Motion language, whether a slow dolly-in, a handheld tremor, or a locked-off static take, tells the model how the scene should feel in time. Use these terms deliberately and the results will read as authored rather than generated.
Negative prompts to purge the artifacts
Even the best generators happily include artifacts you never asked for. Common culprits in photorealistic video include extra or warped fingers and limbs, cloned crowd members, text and logos burned into the frame, plastic-smooth skin, and subject identity drift between shots. A well-maintained negative prompt prevents most of these before they cost you a re-render.
Build a blocking list tuned to your subject. For any human or creature, exclude "extra fingers, warped hands, distorted anatomy." For realism, exclude "cartoon, 3D render, anime, illustration, plastic, airbrushed." For coherence, exclude "morphing, duplicate subject, inconsistent identity." And generally exclude "text, watermark, logo, blurry, low resolution." If a particular failure keeps appearing, it earns a place in the negative prompt. Treat the negative as a living file that grows with your experience and the artifacts you actually encounter.
Controlling motion and camera movement
Photorealism lives and dies on motion physics. People and creatures should move with natural weight; a head turn, a blink, or fabric catching a breeze must read as physical rather than rubbery. To steer this, describe the action concretely and keep the action in the foreground of the prompt. Narrowing the instruction to one clear motion outperforms a sentence that asks for three simultaneous events, because the model has limited coherent energy to spend.
Camera moves need the same specificity. "Slow dolly toward the subject" and "fast whip-pan to an empty hall" produce very different feelings and require different prompt language. For any shot with meaning, say what the camera does, at what pace, and what enters or leaves the frame as a result. If you want a locked, contemplative scene, say so explicitly; a static camera is a choice, not the absence of a choice.
Keeping characters and scenes consistent
For any project longer than a single shot, consistency is the metric that decides whether the result is a film or a card trick. Fix identity with a strong reference image when the tool supports image-to-video, keep a fixed seed across your attempts, and always describe the same subject, costuming, and light setup in identical terms on every prompt. Do not paraphrase your character from shot to shot; the model will render the paraphrase as a different character.
Scene continuity follows the same logic. Reuse identical wording for environment, palette, and weather across your sequence so the world does not shift between cuts. Before locking a scene, generate it at several angles and confirm the identity holds. The time you invest in validating consistency early is returned many times over when you do not have to re-render an entire edit.
Building complex scenes and environments
When a scene needs many elements, large empty worlds, crowds, or detailed production design, scaffold it one layer at a time. Establish the widest shot first to lock geography and scale, then move in for medium shots, always quoting the same environment description. Add detail incrementally rather than asking for everything at once, because an overstuffed prompt splits the model's attention and dilutes every element.
Environmental specificity is your friend. Instead of "a fantasy tavern," try "a dim stone tavern, wooden beams, hanging oil lamps, dust in the air, warm amber light." Each concrete detail gives the model a real constraint to render believably. The more specific your world description, the less likely you are to get generic plates that contradict your close-ups.
A practical production workflow
Production discipline turns prompt skill into shipped work. Start by writing a one-line project brief that states the tone, subject, and cut. Prepare a reference image and lock a seed you can reproduce. Write your master prompt using the block structure, and keep a consistent negative prompt. Then run short test renders at low resolution to validate the core shot before spending a full render, iterating on the prompt until the identity and motion feel right. Finally, render at target resolution, review take by take, and do a light polish pass with color grading and grain before you call it done.
During that sprint, record what worked. Keep a prompt journal where you note the exact prompt, seed, and what you liked or disliked about the take. Over a couple of projects this journal becomes the fastest path to a consistent personal style, and it protects you from forgetting how you achieved a result you loved.
Stylizing without losing the photorealistic base
Photorealism does not mean you cannot have a strong aesthetic. You can push the output toward a particular mood, color palette, or finish while keeping the physics plausible by adding style tokens to the prompt without breaking the realistic constraints. A film, a color grade, a grain level, or a contrast choice refines the look; describing "muted greens, film grain, high contrast" preserves realism while giving the piece a signature. The key is to change finish and color, not the rules of physics the model respects.
When you find a stylization you love, save it as a reusable style block you append to your master prompt. Over time you assemble a small lexicon of cinematic looks, each one a proven template, that you can apply to new scenes and new subjects with confidence. This turns style into an owned asset rather than a happy accident, and it lets every future project inherit the polish you spent time discovering.
Keep your style blocks lightweight and additive so they layer cleanly onto any new subject without forcing you to re-tune the physics. A style block that names a palette, a contrast, and a grain communicates intent in a few words, and it stays compatible with the concrete light-and-camera language your subject block already carries. That modularity is precisely why a shared style travels well across an entire body of work.
An FAQ on photorealistic prompting
Why does my realistic video still look fake? Almost always because of lighting and motion. Flat light and rubbery, over-smooth motion read as synthetic faster than any resolution issue. Fix those first.
How long should my prompt be? Long enough to cover the blocks, but no longer. Packed prompts dilute attention. If a prompt runs long, prioritize lighting, camera, and subject action.
Do I need negative prompts every time? Yes. Defense against common artifacts is worth one saved render per project, and it compounds.
How do I stop characters changing between shots? Lock a reference, fix a seed, and repeat the exact same character and costuming wording on every prompt. Do not paraphrase.
Can I direct camera exactly? Largely yes, with precise camera vocabulary. Separate the camera statement from the subject action so each controls cleanly.
Continual refinement
Runway Gen 3 gives you a cinematic instrument, but the authorship still comes from you. The craft of photorealistic prompting is the craft of thinking like a cinematographer: naming the light, choosing the lens, deciding where the camera looks, and telling one clear piece of physics truth per shot. Those decisions are what make output feel authored rather than generated, and they are entirely transferable to any model you later adopt.
Start with a single, deliberately crafted shot. Get the light and the motion to read as physical, lock a clean negative prompt, and judge the result against a real frame from a film you admire. Then scale to a two-shot scene, then a sequence. Each step builds the same muscle, and before long the model stops being a lottery and becomes a collaborator you direct.
