Launching Line 1.0

1. Model Overview

Line 1.0 is a deterministic drawing-order model used to convert static instructional imagery into animated, hand-drawn explainers. The model is intentionally designed as a hybrid system: classical geometric vision for interpretability, structured sequencing for reading order consistency, and time-aware rendering for controllable motion. This choice gives us predictable outputs across education, product walkthroughs, and technical diagrams where repeatability matters more than one-shot visual randomness.

In practical terms, the model does three things: it identifies drawable structures, it ranks those structures into a human-readable sequence, and it reveals them with temporally stable stroke growth. The output is therefore not just an animated mask; it is a sequence that approximates how a person would progressively draw while explaining. This is especially useful for multi-part scenes where narrative flow and motion timing must stay aligned with accompanying voiceover and beat-level scripting.

We deliberately expose the conceptual model and mathematical framing while withholding private constants, fallback scoring strategies, and robustness heuristics that materially affect production reliability. This keeps the explanation transparent without disclosing sensitive internals.

2. Algorithm Overview

The algorithm starts with spatial normalization. Every source image is mapped into a fixed render canvas, preserving aspect ratio and centering the content. This avoids distortion while ensuring consistent downstream timing and stroke geometry. After normalization, a style-dependent binarization stage produces an ink map. Solid-style rendering prefers globally stable thresholding, while sketch and pencil styles use adaptive local thresholding to preserve subtle line structure under varying local contrast.

Once an ink map is available, connected components are extracted and each component is converted into one or more contour primitives. These primitives are treated as atomic draw candidates. The next stage applies geometric grouping: candidates are clustered into visual lines using center-distance and height-aware rules, then sorted left-to-right within each line. A controlled merge stage combines neighboring candidates into larger semantic units when spacing and alignment imply they should be drawn as one continuous token.

Temporal rendering is then applied per scene segment. Completed units are fully revealed, while the active unit is partially revealed according to local progress and contour-length budget. This produces the marker progression effect associated with whiteboard drawing, while preserving strict ordering constraints needed for legibility.

3. Mathematical Formulation (High-Level)

The resize and normalization stage uses a standard aspect-ratio preserving transform. Let the input size be (w, h) and target canvas be (W, H). The scale factor is:

Background polarity is estimated from mean grayscale intensity. If I(x,y) is grayscale intensity over N pixels, we compute:

This statistic controls inversion behavior for binary conversion in dark-versus-light backgrounds. For solid mode, threshold selection follows a variance-maximization principle (Otsu family objective); for local sketch-like modes, thresholding uses neighborhood-weighted Gaussian statistics.

Line grouping and merge decisions use geometry constraints. A candidate glyph g with centercy_g joins a line with average center cy_l when vertical deviation remains bounded:

Horizontal merge uses gap testing between adjacent candidates:

Segment assignment for multi-part beats uses center-position partitioning:

Rendering progress is computed over segment time with normalized interpolation:

Here, G is the number of drawable units in the segment and L_total is total contour length for the active unit. This formulation ensures smooth intra-unit growth while preserving exact inter-unit ordering.

We intentionally omit private calibration constants, resilience heuristics, and ranking tie-break logic. Those details are part of model hardening and are not published on this page.