Why Water Is the Hardest Thing to Fake
There is a reason water scenes are the terror of every visual effects artist. Water is everywhere in our visual memory — we see it in glasses, rivers, oceans, rain, and tears — and we know, with an accuracy we cannot articulate, when it moves wrong. Light refracts through it, bends around it, and scatters inside it. It flows, splashes, ripples, and settles in ways that physics has spent centuries describing. Fake it poorly, and the audience feels it instantly, even if they cannot say why.
This is why "in water cinematography" — underwater scenes, surface reflections, rain, liquid interactions — has traditionally been one of the most expensive and difficult things to shoot. Real underwater filming requires specialist cameras, divers, safety teams, and controlled environments. Studio tanks cost fortunes. And even then, the results are limited by what can physically be captured.
AI video generation has changed the economics of water. Modern models can simulate refraction, caustics, bubbles, and surface dynamics with surprising fidelity — when you use them correctly. This tutorial walks through the technical foundations of AI water cinematography and the practical workflow for producing believable underwater and liquid scenes.
The Physics You Cannot Ignore
Before touching any tool, it is worth understanding what makes water look like water. Four phenomena matter more than anything else.
Refraction. Light bends when it passes between media of different densities — air to water, water to glass. Underwater objects appear distorted, shifted, and magnified. Get refraction wrong and the scene reads as a cheap overlay; get it right and the image has an immediate physical truth.
Caustics. Light passing through a curved water surface concentrates into patterns of brightness on submerged surfaces — those dancing, net-like patterns you see on the bottom of a pool. Caustics are the single strongest cue that a scene is genuinely underwater.
Surface dynamics. Ripples, waves, and splashes are driven by energy and gravity. Water has weight; it does not wiggle like jelly. Models that cannot simulate inertia produce water that looks like a screensaver.
Scattering and depth. Water absorbs light unevenly. Reds disappear first with depth, then oranges and yellows, leaving blues and greens. Any believable underwater scene has a color story that shifts with distance — this is why underwater shots look blue-green even in bright sunlight.
You do not need to model these phenomena yourself; you need to choose models and prompts that handle them, and to know which ones to check when the output looks wrong.
Choosing the Right Model for Water
Not all video models are equal when it comes to liquids. Some models have visibly stronger physics simulation; others produce water that shimmers unrealistically or lacks weight.
For photorealistic water with good physics, models with strong physical simulation are your first choice. The Runway line — particularly Gen-4 — has a strong reputation for handling physical properties and camera motion, which matters for dynamic water scenes. PixVerse V4.5 is notable for its cinematic lens controls, giving you the ability to direct focus, aperture, and motion blur in a way that sells depth and scale. Kling AI models handle dynamic motion well and are a good option for action-heavy water scenes like waves, splashes, and storms.
For testing and drafts, cheaper models are fine. The key is to run a physics test before you commit: generate a simple scene — a glass of water, a ripple, a splash — and study how the water behaves. If the draft water has the right weight and motion, the model can handle your hero shots. If it looks floaty, switch models.
The Core Workflow for Underwater Scenes
The workflow below assumes you are producing a scene where the camera is underwater or water dominates the frame. It applies whether the scene is a documentary-style reef shot, a sci-fi flooded corridor, or a fantasy mermaid sequence.
Step 1: Define the Water Story
Decide what the water is doing in the scene. Is it calm and translucent, turbulent and murky, or somewhere between? Write the water into the scene description as a character: "slow-drifting silt, soft refracted light, muffled sound, deep blue-green palette." Specificity here does more for realism than any model setting.
Step 2: Lock the Color Story
Choose the depth of your scene and set the palette accordingly. Shallow water: bright teal, strong caustics, visible sun shafts. Mid-depth: blue-green, softened contrast, floating particles. Deep water: near-black blue, minimal color, high contrast from artificial light. Encode this palette in your reference images and your prompts.
Step 3: Generate Reference Stills First
Before video, generate stills of the underwater environment — the seabed, the light shafts, the caustic patterns, the character in the water. These stills serve two purposes: they validate the look, and they become the reference inputs for the video stage. Consistency in water scenes is brutally hard; reference images are your best defense.
Step 4: Keyframe the Action
Underwater motion is slow, weighted, and buoyant. Hair drifts, fabric floats, bodies move with effort. Use keyframe control to set the start and end of each shot, and describe the motion in the prompt with underwater language: "slow drift," "buoyant," "weightless," "gentle current." Most models will honor this direction if you are explicit.
Step 5: Manage Character Consistency
Characters underwater are the hardest consistency problem in AI video, because the environment itself is in motion. Use multi-image fusion: provide two or more consistent portraits of the character, ideally in the water, and reuse them across every shot. Keep the costume description identical. Expect to regenerate more often than in dry scenes; water scenes are where your patience pays off.
Step 6: Handle Micro-Detail
The difference between professional and amateur water is micro-detail: surface ripples, bubbles rising, silt drifting, light dancing on skin. These are exactly the details that generic models skip. Put them in the prompt explicitly: "rising micro-bubbles," "caustic patterns on the seabed," "suspended particles in the light shaft." If your tool supports it, generate a close-up pass of the detail elements and cut them in during editing.
Directing Light Underwater
Lighting is where water cinematography really lives. Underwater, light behaves differently than on land: it scatters, it softens, and it dies with depth. Directing it with AI requires translating that behavior into prompt language.
Backlighting reads beautifully underwater. Describe "strong backlight from the surface, god rays, silty water glowing in the beam." Side light defines forms in murky water. Practical lights — a diver's lamp, a submarine's floodlight — create depth and drama, and they are easy to describe: "cold white beam cutting through blue water."
The most common lighting mistake in AI water scenes is over-lighting: a bright, even, "studio" look that destroys the illusion of depth. Underwater scenes are dark and moody in reality. If your output looks like a swimming pool commercial, deepen the palette, increase contrast, and move the light source into the scene.
Sound and Music for Water Scenes
Audio is half of the illusion, and water has a rich sonic language: bubbles, muffled thumps, distant surface rumble, the strange quiet of deep water. AI audio tools can generate ambience beds and effects, but the craft is in layering. Start with a deep, low ambience; add bubble pops and clicks for texture; keep music sparse and low-passed, as if heard through water. The moment the audio stops selling the underwater world, the visuals lose their credibility regardless of how good they are.
A Worked Example: The Submerged Corridor
To make the workflow concrete, here is a walkthrough of a typical shot: a character swimming through a flooded sci-fi corridor.
The brief is one sentence: a maintenance worker in an orange suit swims through a flooded corridor, emergency lights strobing, silt drifting in the beams. From this, the scene bible notes the depth (mid-depth, so blue-green with visible light shafts), the palette (cold blue with warm orange accents from the suit and emergency lights), and the character sheet (the worker, the suit, the helmet).
Reference stills come first: three images of the character in the suit, two of the corridor from different angles, one of the light behavior. The best stills are promoted to reference inputs. Then the shot is pre-vizzed on a fast model: a ten-second clip checking composition and pacing. The first pre-viz shows the character drifting too fast; the prompt is adjusted with "buoyant, slow drift, weighted movement."
The hero pass runs on a premium model with the reference images and a keyframe pair: a close-up of the character at the first frame, a wide shot of the corridor at the last. The prompt names the light, the silt, the bubbles, and the camera: "slow dolly backward, emergency lights strobing, rising bubbles, drifting silt, blue-green palette with warm accents." After three variants, one is selected and two artifacts are fixed in the edit. Sound is layered: a low ambience, bubbles, and a muffled alarm. The shot is done.
This example is deliberately small. The value of the workflow is that it scales: the same structure, repeated for every shot, is what turns a single good clip into a complete film.
Prompt Library for Water Cinematography
Build your own prompt library as you work. Here are starting points to adapt:
Calm underwater scene: "Underwater wide shot, clear blue-green water, soft sun shafts from the surface, caustics on the sandy floor, slow-drifting silt, gentle current, photorealistic, cinematic."
Character in water: "Full-body shot of [character], hair drifting in the water, [costume], surrounded by rising bubbles, soft refracted light, blue-green palette, slow motion, shallow depth of field."
Stormy surface: "Wide shot of a stormy ocean, massive waves, spray in the air, dark grey sky, dramatic light, cinematic anamorphic."
Rain and reflections: "Close-up of raindrops hitting a puddle, concentric ripples, reflections of city lights, slow motion, shallow depth of field."
Splash moment: "Slow-motion splash, water droplets suspended in air, backlit, high detail, shallow depth of field, cinematic."
The pattern in all of these is the same: name the water behavior, the light behavior, the palette, and the camera language. Leave the rest to the model.
Troubleshooting Water Artifacts
Water warps or shimmers unrealistically. Usually a model problem. Run the physics test and switch to a model with stronger simulation, or shorten the generation — long generations accumulate artifacts in liquids faster than in solids.
Colors look wrong. Revisit your depth palette. If the scene is deep but the colors are bright, the model is ignoring your depth cues. Add "deep water," "low visibility," or "desaturated" to the prompt, and push the reference stills bluer.
Bubbles and caustics are missing. These are micro-details; name them explicitly and consider a dedicated close-up pass.
Characters morph between shots. Reinforce the reference set: more consistent portraits, multi-image fusion, identical costume text. Expect several iterations; water is the stress test for consistency systems.
Motion is too fast. Underwater motion is slow and weighted. Add "buoyant," "slow drift," "weightless" and use keyframes to constrain the movement.
Frequently Asked Questions
Do I need a special model for water scenes? Not a dedicated one, but you need a model with strong physics simulation. Test before committing.
Can AI handle real underwater footage style? Yes. Describe the camera housing, lens behavior, and lighting of real underwater cinematography in your prompts, and reference real footage stills.
How do I make characters look submerged rather than floating in front of a background? The trick is environmental interaction: bubbles, hair drift, costume movement, light shafts, and color grading. A character with no interaction cues looks pasted in.
Is this suitable for commercial work? Yes, if you respect model licenses and quality standards. Underwater and liquid scenes are among the most requested styles for commercials, music videos, and film pre-viz.
What is the fastest way to learn? Pick one scene type — a glass of water is a great start — and study it until you can generate a believable close-up. Then scale up to pools, then oceans.
Final Thoughts
Water has always been the benchmark for visual effects, because audiences carry an internal physics engine that no marketing team can override. AI video generation has brought water cinematography within reach of independent creators, but the craft has not been automated away — it has been relocated into prompt design, reference management, model selection, and sound.
Master the physics, respect the palette, feed the model good references, and treat audio as half the illusion. Do that, and the hardest thing to fake in video becomes something you can create on demand. The tank is closed. The ocean is now a prompt.
