target: ა
Draw between the guide lines (baseline is the lower one). Strokes are captured as (x, y, t) trajectories — timing gives real pen speed, so generated strokes thin out on fast segments. Each extra sample of a letter improves variability: generation morphs between random pairs of your samples and adds smooth correlated noise.
Every knob is a distribution parameter. The pipeline: each Mkhedruli character is rendered
to its own bitmap → random affine (rotation, scale, shear) → composed with kerning jitter and
occasional negative spacing (fake ligatures) along a baseline that drifts as sine + random walk →
the whole line passes through a smooth random displacement field (Gaussian-filtered noise, the
classic elastic-distortion trick) → morphological erode/dilate varies stroke width → an ink model
adds per-glyph darkness variation and edge bleed onto textured paper.
Batch export writes images/*.png + labels.jsonl + params.json —
ready for HTR training. Static page, no server: hostable on GitHub Pages.