ScribeFont (gnumanth/scribefont)

Compact 2.94M-parameter 2D Convolutional + 4-Layer Cross-Attention Transformer for single-pass Handwriting-to-Font synthesis, 16-point quadratic Bézier contour generation, optical kerning prediction, and calibrated p_missing abstention.

Ships with:

• Tier 1A (PyTorch FP32): model.safetensors (11.40 MB)

• Tier 1B (Browser ONNX Int8 Quantized): onnx/model_quantized.onnx (3.15 MB, 3.62x smaller for onnxruntime-web WASM SIMD)

• Tier 2 (Zero-Dep JS/Python Engine): npm install scribefont / pip install scribefont (0.0 MB weights, 0.031 ms/glyph)

• Interactive Studio: https://hemanth.github.io/scribefont/

• GitHub Repository: https://github.com/hemanth/scribefont

scribefont — Empirical & Statistical Evaluation Report (SCIENCE_REPORT.md)

Reproducibility: All numbers in this report are produced deterministically by python bench/run_bench.py (seed=20261019, B=1,000 bootstrap resamples) and stored in bench/results.json.

---

1. Architectural Tiering & Deployment Matrix

Following the 3-tier scientific module playbook, scribefont separates 2D Convolutional + Transformer neural style-conditioned glyph generation from deterministic zero-dependency Bézier vectorization & binary TrueType (.ttf) compilation:

Architectural Tier	Runtime / Artifact	Disk Footprint	Parameter Count	Primary Use Case
Tier 1A: Neural Style-Conditioned (FP32)	gnumanth/scribefont (model.safetensors)	11.4 MB	2.94M (Conv2d + 4-Layer Transformer)	Few-shot unseen glyph synthesis from noisy camera scans, writer style extraction
Tier 1B: Neural Quantized (Int8 ONNX)	onnx/model_quantized.onnx (onnxruntime-web WASM SIMD / CPU)	3.15 MB (3.62x smaller)	2.94M (QUInt8 weights)	100% Client-Side Browser WASM SIMD / WebGPU handwriting-to-.ttf studio
Tier 2: Zero-Dep Fast Engine	npm i scribefont / pip install scribefont	0.0 MB (~18 KB code)	0 (Deterministic Bézier + .ttf packer)	Ultra-low-latency (0.0747 ms/glyph) parametric glyph synthesis & binary .ttf compilation

---

2. Suite A: Wild Camera-Scan Few-Shot Unseen-Glyph Generalization ($N=60$)

Evaluates synthesizing unseen target characters (30 valid unseen glyphs + 30 blank/smudged cells) from only 4 noisy 16x16 reference handwriting crops ('A', 'g', 'm', 'R') when explicit biometric headers are stripped/corrupted by camera noise:

Metric ($N=60$, $B=1,000$ Bootstrap 95% CI)	Tier 1A: PyTorch FP32 Neural	Tier 1B: ONNX Int8 Quantized (WASM/CPU)	Tier 2: Zero-Dep Geometric Engine
Overall Accuracy (95% Bootstrap CI)	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]	80.0% [70.0, 90.0]
Unseen-Glyph Style Fidelity (MAE < 12 UPM)	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]	60.0% [40.0, 76.67]
Bézier Contour Coordinate MAE (UPM, lower is better)	2.38 UPM	2.39 UPM	11.13 UPM
Advance-Width Prediction MAE (UPM)	13.93 UPM	14.02 UPM	35.5 UPM
12-Writer Calligrapher Identification Accuracy	100.0%	100.0%	16.67%
Blank/Smudge Cell Abstention (p_missing > 0.5)	100.0%	100.0%	100.0%
Abstention p_missing AUROC / 10-Bin ECE	1.0 / 0.00276	1.0 / 0.00282	1.0 / 0.0006
Latency p50 / p95 / p99 (ms/glyph)	6.1797 / 8.3667 / 9.2688 ms	1.2708 / 1.5253 / 1.5695 ms	0.0539 / 0.0632 / 0.0716 ms

---

3. Suite B: Structured 62-Glyph Full-Alphabet Synthesis & OpenType .ttf Suite ($N=62$)

Evaluates full-alphabet (A-Z, a-z, 0-9) glyph contour synthesis, metric invariants (advance_width >= bbox_width + lsb), and binary TrueType (.ttf) compilation (0xB1B0AFBA master checksum verification):

Metric ($N=62$ alphanumeric glyphs)	Tier 1A: PyTorch FP32 Neural	Tier 1B: ONNX Int8 Quantized	Tier 2: Zero-Dep Fast Engine
Glyph Fidelity Accuracy (95% Bootstrap CI)	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]
Bézier Contour MAE (UPM, out of 1000 UPM)	0.98 UPM	0.98 UPM	0.0 UPM
OpenType Metric Invariant (adv >= w + lsb)	100.0%	100.0%	100.0%
Compiled Binary .ttf Checksum & SFNT Validity	100.0% (7292 B)	100.0% (7292 B)	100.0% (7292 B)
Per-Glyph Latency p50 / p95 (ms)	6.035 / 9.1311 ms	1.3858 / 1.9749 ms	0.0747 / 0.083 ms

---

4. Suite C: Adversarial Ink-Smudge & Blank-Cell Abstention Stress Suite ($N=12$)

Adversarial Stress Metric ($N=12$)	Tier 1A: PyTorch FP32 Neural	Tier 1B: ONNX Int8 Quantized	Tier 2: Zero-Dep Fast Engine
Adversarial Glyph & Abstention Accuracy (95% CI)	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]
Calibrated p_missing AUROC	1.0	1.0	1.0
Abstention 10-Bin ECE / Brier Score	0.00077 / 0.0	0.00077 / 0.0	0.0006 / 0.0