vectormark-logo-pipeline

vectormark

Deterministic logo idealizer — turn a rendered (raster) logo into a clean, editable, exactly-symmetric SVG.

Where conventional tracers (vtracer, potrace, autotrace) chase pixel edges and emit anonymous <path> soup, vectormark recognizes structure: it collapses anti-aliasing to a true palette, detects symmetry, and fits ideal primitives (<ellipse>, <rect>, regular polygons) to regions — falling back to fitted Bézier paths only where no primitive matches. The output SVG doubles as a human-editable parametric model.

No model/LLM in the loop. Every stage is a known, deterministic algorithm.

Status

🌱 v1 pipeline working. The deterministic A0→B→C pipeline idealizes the Daikonic mark end-to-end (SSIM 0.98 / mean ΔE 0.006 vs. source) into a 1.9 KB structured SVG with exact bilateral symmetry. See the design spec and plan:

• spec: docs/superpowers/specs/2026-06-04-vectormark-logo-idealizer-design.md
• plan: docs/superpowers/plans/2026-06-04-vectormark-v1.md

uv sync --extra dev
uv run vectormark path/to/logo.png -o out.svg
uv run pytest

MCP integration is available for AI clients that can call local stdio servers (the server lives in the optional server extra; scored selection needs the scoring extra — without it, selection falls back to the cascade pick):

uv sync --extra server --extra scoring
npm --prefix integrations/mcp-app install
npm --prefix integrations/mcp-app run build
uv run vectormark-mcp

See docs/mcp.md for client configuration and hosted app notes.

Pipeline (v1)

A0  color optimization   robust palette (OKLab, ΔE-merge); AA halo discarded as noise
B   trace + clean        vtracer per layer; drop background plate
C   idealize             symmetry axis (PCA + reflection gate) → fundamental domain
                         → contour → RDP → primitive-first / Bézier-fallback
                         → geometry-level mirror (<use>)
→   structured SVG       native shapes + <use>; optional --flatten to paths

v1 targets flat-color segmented marks. v1.1 adds gradient normalize/re-apply.

Prototype → core

Built first in Python (numpy/scipy/scikit-image/shapely/Pillow + vtracer), with a pure, IO-free pipeline so the eventual Rust core (npm via napi-rs, PyPI via PyO3/maturin, WASM for browser) is a mechanical port.

License

MIT