Algorithmic film — ongoing
No footage. No actors. No authored meaning.
Each run produces something unrepeatable.
The algorithm generates. I curate. I select from what the process produces — not what I intended, because I intended nothing. My direction is selection from process, not generation of content. What results is film that had no author until the moment of choice.
Reaction-Diffusion · Gray-Scott model
Two virtual chemical species compete across a grid. Their interaction produces complex, spontaneously ordered patterns: spots, labyrinths, waves. No instruction. No plan. Only competition.
Alan Turing proposed in 1952 that biological form — leopard spots, zebra stripes, the arrangement of fingers — emerges from this mechanism. The implication: biological complexity is not designed.
In developmentAbelian Sandpile · Per Bak model
Grains fall. The pile builds toward a threshold it will always exceed. Long stillness. Unpredictable cascades. The system self-organises to the edge of collapse — and stays there.
Per Bak claimed this single model explains earthquakes, extinctions, market crashes, and neural activity. The implication: most complex systems are always at the edge. Stability is not a stable state.
In developmentPhysarum Polycephalum · Slime mold
A model of slime mold growth. Branching paths form, thicken with use, thin with disuse, dissolve when abandoned. Sophisticated problem-solving without a brain, a plan, or a centre.
Physarum reconstructed Tokyo's rail network — optimally — using only its starting conditions. The question: what is intelligence if a mold without neurons can do this?
In developmentStrange Attractors · Lorenz, Rössler, Clifford
Deterministic equations trace trajectories that never repeat and never escape. A path accumulates. The centre is never reached. The orbit is the whole.
The Lorenz equations are fully deterministic — given exact starting conditions, the future is calculable. In practice, unmeasurably small differences produce completely different behaviour. Determinism and predictability are not the same thing.
In developmentCellular Automata · Conway, Brian's Brain
A grid of cells. Each cell: on or off. Each cell's future determined by its immediate neighbours. From three local values: complexity, propagating structures, apparent purpose.
Rule 110 is computationally universal — it can simulate any computation, including itself. Stephen Wolfram argued that simple computational rules could explain most phenomena in physics. The claim is still live.
In developmentDiffusion-Limited Aggregation
Particles random-walk until they meet the growing cluster and stop. The structure branches, becomes its own obstacle. Growth happens only at the tips. The interior empties.
The same algorithm models phenomena across 20 orders of magnitude of scale — from electrodeposition of metal crystals to the large-scale structure of the universe. The process has no sense of scale.
In developmentInformation Entropy · Shannon / Boltzmann
Visual fields at calculated entropy values. The transit between structure and noise rendered directly. Order borrowed. Noise destination.
The second law of thermodynamics says entropy always increases. Life is a local, temporary exception. Intelligence is a local, temporary exception. The question of why the universe started ordered remains unresolved.
In developmentStep 01
Parameter space
Each algorithm exposes a set of controls: feed rate, kill rate, sensor angle, stickiness, entropy level. The parameter space is the only place human decision enters the generative process.
Step 02
Generative run
The algorithm runs on blank or noise data for the specified duration. Frames are written continuously. No footage. No models trained on existing images. The output exists nowhere until it is generated.
Step 03
Selection
From many runs, a small number are selected. The criterion is not definable in advance. There is something in certain runs that is absent in others. This is where the director's work lives.
Step 04
Sequencing
Selected runs are cut and ordered. Duration is deliberate. The score is commissioned after the edit — written to the image, not the image cut to music. Sound does the emotional labour the image refuses to do.
Step 05
Output
24fps. 4K. Films between 10 and 60 minutes. Single-channel for festivals. Multi-channel for installation. All rendered offline. No cloud. No GPU required.
Step 06
Distribution
Festival submissions. Gallery installation. Community screening. Streaming. Each context shapes what the work is. The algorithm is indifferent to venue.
Stack
Python 3.10 · NumPy · SciPy · OpenCV · FFmpeg
No neural networks. No training data. No API calls.
Output specifications
Resolution: 4K (3840×2160) or 1080p (1920×1080)
Frame rate: 24fps
Duration: 10 – 60 minutes per programme
Format: H.264 / ProRes for installation
Render time (1080p, 10 minutes)
Reaction-Diffusion: ~90 min
Slime Mold: ~60 min
Strange Attractor: ~45 min
Cellular Automata: ~30 min
All others: ~20–40 min
Each run is unique
Seed values produce reproducible outputs, but no two editorial selections are identical.
A work screened twice is the same file. A work regenerated is a different film.
Natascha works with algorithms that produce visual material from blank or random data. She does not direct the algorithm. She selects from its output.
Her interest is in the gap between process and apparent intention. These systems do not try to produce anything. They operate according to local rules. The results look almost purposeful. That gap — between mechanism and meaning — is where the films live.
She is credited as director. What this means: she decided what to show, for how long, in what order, with what score. The algorithm decided everything else.
She does not argue that this is art. The argument is the work's problem, not hers.
Response within one week.