[ Audiovisual research / 003 ]
LOTHAL LAB · SANTIAGO ·
Continuous
River.
Mass correspondence and fluid simulation for generative video.
- Type
- Applied research
- Author
- Lothal Lab
- Published
- Updated
- Output
- 3 audiovisual pieces
- Engine
- Python / local
- Version
- Document 1.0
Summary
Continuous River documents two studies and three generative videos. The first transforms two painting series—ten Monet and ten Van Gogh works—without cuts, fades or frame-by-frame editing. The second creates marbled ink through a 2D simulation without using a painting as reference.
Each piece is generated with custom software from recorded parameters and seeds. Rendering runs locally and uses no cloud generative services.
Questions
- 01
How can a painting sequence transform without stopping as it reaches each image?
- 02
Can a 2D simulation create marbled ink without using a reference image?
Short answer
Transform without stopping the motion.
We represent every painting as color masses, solve correspondences between states and interpolate shape, pigment and movement within a global clock. The sequence crosses all ten works and returns to the beginning without pausing.
Generate ink from blank paper.
A stable-fluid simulation transports ink and vorticity events until marbling emerges. The process uses no painting or other reference image.
Study I / Pictorial flow
Representation through color masses.
The corpus contains ten paintings by Claude Monet and ten by Vincent van Gogh, from public-domain reproductions available on Wikimedia Commons. Both series use the CAUCE engine’s meandro_fspline configuration.
Series order
The paintings are ordered to maximize color affinity between consecutive states. With ten images, the sequence is solved exactly through Held–Karp dynamic programming.
Decomposition
Mean shift reduces local variation. K-means groups pixels by color and position with k=13; watershed refines boundaries toward image gradients.
Correspondence
A cost matrix combines color, area and position. The Hungarian method solves assignments; the lineage supports appearances, disappearances, splits and merges.
World lines
Correspondences are chained into one trajectory per mass. The meander planner penalizes turns, abrupt speed changes and unnecessary travel.
Shape and pigment
Contours are interpolated through signed distance fields. Color uses a subtractive Kubelka–Munk approximation instead of additive RGB fading.
Continuity and closure
A curl-noise field sustains rotational movement. One global clock joins every segment and returns the final state to the first.
Published videos
Results.
The videos are hosted on media.ymo.cl and load when playback begins.
Ten Monet
Ten paintings · 40 s · 1350 × 1080 · 60 fps · seamless loop · silent
The engine processes soft gradients and diffuse boundaries between color regions.
Ten Van Gogh
Ten paintings · 40 s · 1350 × 1080 · 60 fps · seamless loop · silent
The same configuration is applied to works with stronger contrast, impasto and brush direction.
Marble
Three-panel polyptych · 90 s · 1080 × 1350 · 30 fps · silent
A marbling simulation inspired by ebru, with no pictorial reference image.
Study II / Marble
Simulation and temporal sequence.
Marble begins with blank paper and a fixed palette. Each panel contains a 2D simulation based on Stam’s stable-fluid method. Ink is represented as absorbance and transported through a limited MacCormack scheme. The engine uses Python, NumPy and OpenCV; the published render ran on CPU with seed 23.
Paper
No ink is injected during the first 1.5 seconds.
Accumulation
The event rate increases and the delay between panels decreases.
Climax
A radial pulse and increased vorticity redistribute the pigment.
Rest
Velocity is damped. After freezing, only density and saturation change.
Verification
Continuity criterion.
The judge calculates the mean absolute difference between each pair of frames. A render is rejected when the minimum value around a transition is below 0.5 times the clip median. The test detects accidental pauses; it does not evaluate visual quality.
This is an internal project metric, not an external standard.
Design decisions
The corpus selection, correspondences, trajectories and rejection criteria were defined during development.
These decisions are recorded in the code and its parameters, rather than added through later video editing.
Scope and limitations
Study limitations.
- The metrics verify continuity and boundary fit; they do not measure artistic quality or audience response.
- The temporal-change threshold is an operational project criterion, not an external standard.
- Kubelka–Munk is used as a subtractive approximation without spectral calibration against physical pigments.
- Marble is a fluid-inspired visual simulation and has not been validated against a physical ebru experiment.
- Exact reproduction requires the same seed, dependency versions and compute path.
- The individual URL for each Wikimedia Commons reproduction remains to be recorded.
Painting corpus
Works used.
Claude Monet
Impression, Sunrise · The Japanese Bridge · Poppies · The Magpie · Haystack in Sunlight · Houses of Parliament · San Giorgio Maggiore · Terrace at Sainte-Adresse · Cliff Walk at Pourville · The Saint-Lazare Station.
Vincent van Gogh
The Starry Night · Starry Night Over the Rhône · The Harvest · Fishing Boats · Irises · Almond Blossom · Wheatfield with Crows · The Bedroom · Wheat Field with Cypresses · The Red Vineyard.
Reproductions gathered from Wikimedia Commons. Titles are normalized in English for this document.
References
Technical references.
- 01
Held, M. and Karp, R. M. (1962). A Dynamic Programming Approach to Sequencing Problems.
- 02
Kuhn, H. W. (1955). The Hungarian Method for the Assignment Problem.
- 03
Kubelka, P. and Munk, F. (1931). An Article on Optics of Paint Layers.
- 04
Stam, J. (1999). Stable Fluids.
- 05
Bridson, R., Houriham, J. and Nordenstam, M. (2007). Curl-Noise for Procedural Fluid Flow.