A real-time, physically based ocean in pure WebGL2, with weather you can steer: calm golden hours, rising wind, rain, squall lines, gales, lightning and moonlit nights. There are no textures, models or image assets. Every wave, cloud, foam bubble and raindrop is computed. The only recorded material is sound (rain and thunder).

Live: https://leonxlnx.github.io/open-water/ · Film: media/open-water.mp4 (one minute, 1080p60, rendered offline by the same engine; the full-quality master is on the releases page).

The film runs in three acts: a golden-hour calm; a storm arriving as the sky clouds over, from building cumulus to a low, dark deck, then rain, a gale and lightning; and a moonlit night after the storm, with the cell still flickering on the horizon.

Waves. Tessendorf FFT synthesis on the GPU with four cascades (1800 m down to 3.7 m, so a gale's ~300 m peak is carried by several modes and forms wave groups) and capillary-gravity dispersion. The wind sea uses the Elfouhaily et al. (1997) unified spectrum, whose short-wave regime is calibrated to Cox–Munk slope statistics. Spreading is Donelan–Banner near the peak and ECKV for short waves, and, for the shorter waves, blended toward a broad cos²ˢ spread as the wind approaches a gale (a confused storm sea under long-crested dominant waves), plus a narrow-band swell. Choppy displacement gives trochoidal crests, steeper as the wind rises (young storm seas). Significant wave heights match the Beaufort table (Bf 5: 2.1 m, Bf 7: 4.8 m, Bf 9: 9.2 m), and the GPU statistics are checked against the spectrum (tools/hscheck.mjs).

Light on water. A slope-variance (LEAN) BRDF after Bruneton, Neyret and Holzschuch (2010) makes geometry, normals and BRDF hand over seamlessly. Sub-cascade roughness is the Cox–Munk remainder, and no pixel may hold more slope variance than the whole surface has. Sun glitter is carried by individual capillary facets: a pixel footprint holding N facets shows on average λ = E[L]·Ω☉ / (F·L☉)·N glints, sampled as a Poisson count, so the near field is discrete, full-brightness sparkles on dark water and the far field converges to the smooth glitter path. Near the camera the glinting facets are larger, so sparkles are fewer and brighter; each one's brightness varies, lasts a few frames, and at grazing angles is drawn out along the foreshortened wave. A rough sea blurs the reflected sky mostly in elevation: tilting a facet sideways turns a grazing reflection by only about el·tan(e) in azimuth. A ray reflected below the horizon sees the next wave. Also modeled: averaged Fresnel, water-leaving radiance from ocean-optics remote-sensing reflectance (deep ultramarine that turns grey-green as breaking waves fill the water with bubbles), backlit crest scattering, cloud shadows drifting over the sea, Earth curvature and aerial perspective.

Foam. Whitecaps form where the surface Jacobian folds. Two stages are tracked: active breaking (dense froth that lives about 0.4 s) and residual foam (seconds to tens of seconds). Both are drawn as the probability that a procedural bubble-raft pattern lies below the local foam density, so a whitecap is solid only at its core, its rim frays into filaments, and old foam opens up along the raft network into lace instead of shrinking as a blob. Crests break in segments of a few metres along the crest. Residual foam lives from about a second in a breeze to twenty seconds in a gale and fades out as it thins, and at gale force gathers into narrow, meandering wind rows tens of metres long. Fresh caps have an opaque white core; foam is a translucent layer over the water, never darker than it. The breaking threshold is expressed in standard deviations of the Jacobian and calibrated so coverage follows the Monahan & O'Muircheartaigh (1980) whitecap law within about 20% from 8 to 21 m/s, with active breaking about 35-55% of it (tools/foam.mjs).

Sky. The atmosphere follows Hillaire (2020): transmittance, multiple-scattering, sky-view and aerial-perspective LUTs. Clouds are volumetric, built from procedural Perlin–Worley noise with Beer–powder lighting, Wrenninge multiple-scattering octaves and two-stream diffusion through thick storm decks (a deck transmits E·μ₀·T/π). Faces turned to the camera reflect the sky and sea light behind it, so a storm's shadowed side is dark blue-grey, not black and not glowing. The marine aerosol scatters with a three-lobe phase function (a sharp core inside a graded aureole, a little backscatter) and has a realistic load (visibility around 100 km in clean air, a few tens of km in a storm), which gives the solar aureole, hazy golden horizons and distant clouds fading out. The cloud pass also records how far away its light comes from, so rain and mist are drawn in front of a cloud only up to that distance.

A squall line moves the weather across the sea along a meandering front. It is a row of cumulonimbus cells about 10 km tall whose anvil overhangs ahead of the line; its face is displaced by kilometre-scale 3D noise that grows with height, so the towers bulge and recede. Haze standing in the storm's shadow scatters no direct sunlight, and neither do its rain curtains. Scud races beneath the storm near the camera. Storm clouds are optically thick like real cumulonimbus (shaded faces at about half the sky's radiance), and the rain curtains carry broad, uneven shafts that reach up into the cloud base. A storm deck is mottled from below, its optical thickness varying two- to threefold over a few kilometres, and rain falls in drifting cells a few kilometres across rather than as one uniform sheet, so light rain leaves gaps that show the deck and the sea beyond. Crepuscular rays are traced through the haze using the cloud-shadow map. Spindrift is torn off actively breaking crests at gale force and drawn out along the wind, and spray haze hugs the sea.

Rain and storm. Drop sizes follow Marshall–Palmer, with terminal velocities from Atlas et al. (1973). Streak photometry follows Garg & Nayar: a drop covers its streak only D/(v·T) of the exposure. Ring waves expand at the capillary minimum speed, every impact leaves a centimetre-wide crater that relaxes in about 60 ms (some 2,600 per square metre and second in the heaviest rain), and splashes are crowns a few millimetres tall that last about 50 ms; only the largest drops throw 2–4 cm Worthington jets. Rain curtains have shafts, and visibility falls with rain rate (Atlas: β = 0.336·10⁻³ R^0.63 per metre). Rain barely absorbs, so a layer several optical depths thick is lit by the light entering it from the cloud deck and scatters it many times: rain greys the view without blacking out the day, and no direct sun reaches it under a closed deck. Lightning has multiple return strokes, a tortuous fractal channel whose branches taper and dim, and in-cloud illumination. It lights the scene with I/d² (about 100 lux at 1 km) relative to sun- or moonlight, so a daytime strike barely brightens the sea, while at night it outshines the moon. The channel is hidden inside the cloud and dimmed by the rain between it and the viewer. Thunder arrives at 343 m/s and loses its high frequencies with distance.

Night. The full moon takes over as the light source after civil twilight. This gives a moon glitter path, moonlit clouds and a halo in the haze. Stars are point sources with physical flux: the magnitude distribution N(<m) ∝ 10^(0.5m), colour temperatures, extinction, and scintillation near the horizon. They show only once the sky is dark enough. A scotopic (Purkinje) shift drains colour and tints the low light blue.

Camera and image. The camera floats on the sea, heaving, pitching and rolling with waves longer than a small boat. Exposure meters the incident light, but meters the bright sky when shooting into a low sun, as a photographer would, and storms are exposed dark. White balance is near daylight, so golden hour stays gold; under a closed overcast it partly neutralises the cool light within a camera's AWB range. AgX tone mapping is followed by a photographic S-curve, plus a bloom with geometrically falling level weights (a tight lens core, a faint wide skirt) and film grain.

Sound. Sea wash, whitecap hiss and gusting wind are synthesized from the same weather state; the noise beds loop seamlessly and water slaps against the hull as soft, low thuds. Rain and thunder are recordings. The film's soundtrack is rendered by the same engine in an OfflineAudioContext.

Every visual decision was compared against real photographs, side by side (tools/compare.mjs). The references were the NOAA Beaufort-scale series, open-ocean, overcast, rain-squall and storm photography. The measurable parts are checked with tools:

Side by side with NOAA's Beaufort-scale photographs. The render is on the left, seen from a ship's deck 14 m up. The photos are from the NOAA Photo Library (public domain) and carry their film stock's colour cast. At force 9 the photo's sea is darker and its foam lace finer than the render's.

Force 5 (8.5 m/s)	Force 7 (15 m/s)	Force 9 (21 m/s)

A multi-agent review compared 13 scenes with real photographs and had independent judges confirm each discrepancy. A second adversarial code review then confirmed 30 bugs, every one fixed. In later rounds two critics per scene (a photographer and an ocean and atmosphere physicist) scored the renders against some 60 reference photographs (Wikimedia Commons, NOAA), a verifier confirmed each issue, and a regression judge compared every scene with the previous round. Problems found and fixed this way included:

• Rain blacking out the day: the rain was lit by its own dimmed light, leaving only warm stray sun.
• Khaki, mottled storm seas: a rough sea averaged the sun's aureole into every sky reflection.
• Cloud history lagging behind the film's time-lapse squall front as feathered echo bands.
• Contour bands on storm bases from ray-march samples at identical depths.
• A cloud march that ran out of steps before reaching a storm 20 km away.
• The storm sea's directional spreading computed but never applied.
• Sun glitter rendered as a milky sheet instead of sparkles.
• Whitecaps missing at Bf 5.
• Overcast decks without texture (a flat base and too little thickness variation).
• A brown cast from sunlit haze under a storm deck.
• Hairline crest artifacts from mirrored downward reflections.
• Foam outlines with polygonal edges from thresholding a bilinear field.
• Waves travelling against the wind (sign of the time evolution).
• A NaN sea below 2.2 m/s wind (negative Elfouhaily short-wave coefficient).
• A veiled, compressed tonal range (highlights never above ~190/255).
• Whitecaps unrelated to the dominant crests.
• A seam in the sky reflection.
• Lightning washing the whole frame out in daylight.

URL parameters: ?preset=storm, any weather key (for example ?wind=14&rain=0.4), elev and az for the sun, h for eye height, and q=high for film quality.

npm run dev

Then open http://localhost:5173. No build step and no runtime dependencies. You need a browser with WebGL2 and EXT_color_buffer_float, which covers every current desktop browser. The default quality tier targets integrated GPUs; add ?q=high for the film settings.

npm install
npm run dev
npm run render -- --fps 60 --w 1920 --h 1080 --out out/open-water.mp4

The renderer drives headless Chrome deterministically, frame by frame. Frames are piped into ffmpeg/x264 and the soundtrack is muxed in. The timeline is in src/film.js.

src/ocean/   FFT cascades, spectrum, foam, surface shading, surface probe
src/sky/     atmosphere LUTs, volumetric clouds, sky and reflection probe
src/fx/      rain, ripples, spindrift, lightning, post-processing
src/audio/   sound engine (live and offline)
src/film.js  the film timeline
tools/       dev server, screenshots, calibration (wave heights, foam coverage), profiling
render/      offline film renderer

• J. Tessendorf, Simulating Ocean Water, SIGGRAPH course notes, 2001.
• T. Elfouhaily, B. Chapron, K. Katsaros, D. Vandemark, A unified directional spectrum for long and short wind-driven waves, JGR 102(C7), 1997.
• M. Donelan, J. Hamilton, W. Hui, Directional spectra of wind-generated waves, 1985.
• E. Bruneton, F. Neyret, N. Holzschuch, Real-time Realistic Ocean Lighting using Seamless Transitions from Geometry to BRDF, Computer Graphics Forum, 2010.
• C. Cox, W. Munk, Measurement of the Roughness of the Sea Surface from Photographs of the Sun's Glitter, JOSA, 1954.
• E. Monahan, I. O'Muircheartaigh, Optimal power-law description of oceanic whitecap coverage dependence on wind speed, JPO, 1980.
• S. Hillaire, A Scalable and Production Ready Sky and Atmosphere Rendering Technique, EGSR 2020.
• A. Schneider, The Real-time Volumetric Cloudscapes of Horizon Zero Dawn, SIGGRAPH 2015.
• M. Wrenninge et al., Oz: The Great and Volumetric, SIGGRAPH 2013 talk.
• K. Garg, S. Nayar, Photorealistic Rendering of Rain Streaks, SIGGRAPH 2006.
• J. Marshall, W. Palmer, The distribution of raindrops with size, 1948.
• T. Sobotka, AgX display transform.

The code is MIT licensed. The audio recordings keep their Creative Commons licenses; see assets/audio/CREDITS.md.