Concept. Type is usually read, not seen. Kinetic Type strips away legibility and asks the viewer to experience letterforms as physical objects — weight, surface, momentum. Each character was modelled in 3D and assigned unique material properties before being released into a simulated environment.

Early wireframe tests showing the letter “K” built from 14,000 vertices, each vertex responding to gravity and wind vectors.

Process. Built with Three.js and a custom Verlet physics engine. Each typeface was dissected into individual glyphs, re-extruded as 3D geometry, and animated through a system of constraints — spring forces, collision boundaries, and turbulence fields. The music track was generated from the vertex positions via MIDI mapping.

Approach. Rather than keyframing shape transitions by hand, Fluid Forms uses a 2D Navier-Stokes solver to drive the deformation of geometric primitives. Velocity fields from the simulation push vertices across the canvas, creating organic morphing sequences that never repeat.

Technical notes. The simulation runs on GPU via WebGL compute shaders. Density, velocity, and temperature fields are resolved at 120 iterations per frame, then mapped to vertex displacement on a torus primitive. Colour palettes are sampled from the fluid temperature gradient.

Still from the final render — a single frame of the fluid-driven torus displacement.

In-engine screenshot of the terrain system — elevation mapped from layered Perlin and Worley noise.

System design. The landscape is entirely procedural: no hand-placed objects, no pre-baked textures. A multi-octave noise stack generates elevation, moisture, and temperature maps at runtime. Vegetation distribution follows ecological rules — conifers at high elevation, deciduous in valleys, sparse coverage on arid slopes.

Interaction. Viewers navigate with WASD + mouse, or let the autonomous camera tour the terrain. A day/night cycle runs in 8-minute loops, with atmospheric scattering shifting from golden hour to deep blue twilight. The entire stack runs at 60fps on consumer GPUs.

Core idea. What does a camera see when it stops trying to be accurate? Neural Drift feeds live webcam frames through a lightweight style-transfer network trained on early 20th-century landscape painters. The output is a real-time oil-on-canvas abstraction of whatever the camera points at.

Before and after — raw camera input (left) processed through the style-transfer model (right).

Architecture. The model is a MobileNet-v2 encoder paired with a custom decoder, quantised to run client-side via TensorFlow.js at 12fps on a laptop. Latent-space interpolation between style references creates continuous drift even when the camera input is static, giving the piece its meditative quality.

Frame from the opening sequence — Copan building at golden hour, shot on Kodak Vision3 500T.

Background. São Paulo is a city of extreme contrasts: glass towers next to crumbling concrete, silent courtyards steps from roaring avenues. Urban Pulse follows three characters — a street vendor, a bus driver, and a graffiti artist — through their daily routines, using the city’s architecture as a silent fourth protagonist.

Technical. Shot on an Arri SR3 with 16mm Kodak stock. No stabilisation — the handheld camera work is intentional, mirroring the city’s unsteady rhythm. Sound was recorded binaurally and layered with a sparse electronic score composed from field recordings of subway trains and construction sites.

Premise. Still / Ness is a study in patience. Each shot is a single fixed-frame take lasting between 4 and 12 minutes, capturing the slowest visible transitions in coastal light. The title is a double meaning — “stillness” as in absence of motion, and “still” as in the photographic frame.

The camera locked on a tide pool for 11 minutes — watching light creep across barnacles and kelp.

Exhibition. Originally shown as a three-channel installation at a gallery in Lisbon. Each projector displayed a different coastline — Portugal, Brazil, and Japan — creating a triptych of simultaneous dawns. The only sound was the ambient room noise, letting each viewer’s presence become part of the piece.

Process. Every piece begins as a text prompt, but never the first one. The method involves writing dozens of variations, feeding them through a fine-tuned diffusion model, then selecting, blending, and re-prompting based on what emerges. The machine proposes; the human curates.

Tooling. Built with Stable Diffusion XL and a custom LoRA trained on the artist’s own photographic archive. This creates a visual signature that persists across generations — a consistent texture palette, lighting sensibility, and compositional grammar that makes each piece feel like it belongs to the same hand, even when the subject changes completely.

Detail view showing the texture granularity preserved through the LoRA fine-tuning process. The model learned to reproduce film grain and lens flare from the training set.

Premise. What does it mean to look at a face that has never existed? GAN Pulse streams a continuous feed of AI-generated portraits at 24fps, each one unique. The piece runs on a latent-space random walk, ensuring smooth morphological transitions between subjects — a face ages, changes gender, shifts ethnicity, and morphs into the next identity without cuts.

Architecture. StyleGAN3 backbone with a custom latent-space interpolator written in CUDA. The random walk uses Perlin noise in 512-dimensional space to create organic transitions. A separate classifier network filters out low-quality generations in real time, ensuring only convincing portraits reach the screen.

System. Brand Atlas is not a logo — it’s a language. A modular grid system, a set of data-visualisation components, a colour palette derived from scientific imaging, and a typographic hierarchy built around legibility at every scale. The identity lives in the system, not in any single mark.

The grid system in use — every layout, from business card to billboard, snaps to the same 8-column modular grid. The only variable is scale.

Typography. Custom typeface based on DIN Pro, modified for improved readability at small sizes on screen. The number set was completely redrawn to align with the studio’s data-heavy output — each numeral fits precisely within the grid unit.

Approach. Most creative software buries the canvas under toolbars. These studies start from the opposite premise: the canvas is everything. Tools appear only when summoned by gesture, and disappear when not needed. The interface becomes a layer of the creative process, not a separate application to learn.

Motion. Every transition is designed to feel physical — spring animations, overshoot, and inertia give the interface a tactile quality. Buttons compress like real switches, panels slide with momentum, and every action has a micro-feedback loop that confirms the result without text.