AIFlow Math Ink 0.5 is a model-agnostic geometric gridding layer for online
handwritten mathematics. It converts ordered pen strokes into spatially coherent
formula cells and renders one clean image per cell.
This repository contains no OCR model, model weight, tokenizer, training data,
or generated prediction. It is a deterministic preprocessing layer only.
The output images can be passed to any image-to-LaTeX recognizer. They are
interface-compatible with TexTeller-style image batches, but TexTeller is not
included, redistributed, modified, or required by this repository.
AIFlow Math Ink 0.5 pipeline
Why this layer exists
An image-to-LaTeX model normally expects one coherent expression image. A pen
canvas may instead contain multiple rows, detached superscripts, fractions, or
matrix delimiters. AIFlow Math Ink 0.5 uses the original stroke geometry to
partition that canvas before recognition.
The layer:
validates bounded UTF-8 JSON-like stroke events;
computes width-aware stroke bounding boxes;
estimates a scale from the median stroke height;
creates same-baseline and local-attachment edges;
adds structural bridge edges for fraction bars and tall delimiters;
extracts connected components with a disjoint-set structure;
sorts cells in stable two-dimensional reading order; and
re-renders only the member strokes of each cell.
This is a transparent heuristic algorithm, not a trained neural-network layer.
Installation
pip install Pillow
Clone this repository and add src to your Python path, or install it locally:
For an existing API payload, use parse_writing_events() before
build_formula_grids(). See examples/basic.py.
Recognizer compatibility
render_grid_images() returns ordinary RGB PIL.Image.Image objects. A
recognizer can consume the list as a batch without importing this package into
its model implementation. TexTeller compatibility means only this image-level
interface; no TexTeller component is included.
Method and complexity
The complete method, equations, assumptions, and limitations are documented in
docs/METHOD.md.
For (n) accepted strokes, pair construction is (O(n^2)), disjoint-set
operations are effectively near-linear, and rendering is proportional to the
number of retained points plus output pixels. Input caps are mandatory in
server environments; the parser defaults to 2,000 strokes and 200,000 total
points.
Scope and limitations
This release performs geometry-based grouping, not symbol recognition.
It does not infer LaTeX, semantics, or correctness.
Thresholds are scale-adaptive but remain heuristic.
Dense overlapping notes and unusual layouts can over-merge.
The method has not been presented here as a peer-reviewed contribution.
The repository ships no dataset and makes no benchmark claim.
License
The original code in this repository is licensed under Apache License 2.0.
Pillow is an install-time dependency and is not vendored. External recognizers
are separate works governed by their own licenses.
