Exp. 04 / Video Vectorization anantg.xyz →

Temporal Vector Decomposition.

Synthesizing continuous geometric abstractions of cinematic water dynamics at 30 frames per second using sequential evolutionary hill-climbing.

Convex Polygons • 400 Geometric Facets
1920 × 880 • 30.0 fps • 997 sequential frames • ΔRMSE < 0.0001 prune

Sequential Temporal Warm-Starting

Rather than optimizing each video frame independently (which causes high-frequency visual flicker), frame t warm-starts from the exact geometry of frame t - 1. Each vertex is gently guided through local gradient moves, maintaining physical coherence across fluid boundaries.

Pure SVG Vector Representation

All frames are preserved as resolution-independent XML vector shapes with standard viewBox="0 0 1920 880". This allows the animation to be projected at any scale (4K, 8K, architectural installations) without interpolation artifacts.

Adaptive Shape Pruning

The optimizer terminates candidate generation if subsequent shape proposals fail to reduce image error by at least 0.0001 RMSE. This prevents needle artifacts and concentrates polygon density strictly where turbulent motion demands it.