NIV: Neural Axis Variations for Variable Font Generation

NIV (Neural Axis Variations) introduces a neural method that automatically converts static fonts into fully functional variable fonts by predicting per-point displacements conditioned on desired design axes. This work is significant because it automates the labor-intensive process of creating variable font data, allowing for continuous interpolation between styles within a single font file, thereby democratizing typography and enabling flexible design in responsive applications.

Model Architecture and Input Representation

The NIV model employs a sequence-to-sequence geometric model that maps an ordered sequence of glyph point features to a sequence of outline displacement vectors. The input for each glyph is represented as an ordered sequence of points, where each point carries features such as its normalized (x, y) coordinates, an on-curve/off-curve indicator distinguishing between control points and phantom points, and a normalized contour position parameter. To handle the geometric nature of glyph outlines efficiently, the model uses a stack of identical interaction blocks that process the entire sequence simultaneously rather than autoregressively.

Property Embedding Mechanism

A central contribution is the novel Property Embedding mechanism designed to condition the model on multiple design axes simultaneously. This mechanism supports a variable number of axes per example by containing learned vectors for every axis, resulting in a conditioning vector that is weighted by the product of absolute axis values and then normalized. This conditioning vector is used via Adaptive Layer Normalization (AdaLN) at every interaction block to generate the scale and shift parameters, enabling fine-grained, geometry-aware control of the deformation process while maintaining stable training dynamics.

Training Methodology

The model is trained in a supervised manner using ground-truth outline differences computed between the default glyph and a deformed glyph at specified axis configurations. The training data is constructed from existing variable Google Fonts, comprising over one million variation tuples. The loss function applied is the mean squared error applied independently to all points. To ensure strong generalization, the authors utilized different split strategies:

1. Unicode split: Evaluating the model’s ability to generalize deformation patterns to unseen code points based solely on geometric structure rather than character identity.

2. Font split: Measuring generalization across unseen font styles, where entire fonts are assigned to either train or test sets.

Font Generation Process

The second stage involves using the trained model to construct a variable font from a non-variable font in two steps:

1. First, select the desired axes and build the fvar table. The model then predicts outline deltas for each position in the design space, starting with one axis at a time and progressively building up to higher-order sub-spaces.

2. Higher-order tuples are written as residual corrections, where previously inserted lower-order tuples are accounted for using the OpenType tuple-weighting rule (A.4) before subtracting their contribution from the model’s predicted displacement, ensuring accurate multi-axis interpolation.

Generalization and Evaluation

Experiments demonstrate strong generalization across several challenging conditions:

(Unicode split)

(Font split)

The model is also tested on high-complexity CJK glyphs (Meiryo and PingFang) and out-of-distribution handwriting inputs, showing the ability to infer axis-driven deformations for glyph shapes it has never encountered during training. Furthermore, the method outperforms rule-based geometric baselines, which fail to capture semantic deformation behavior. The evaluation metric is the root mean squared error (RMSE) calculated over all test examples. Ablation studies confirm that Property Embedding combined with AdaLN yields the lowest test RMSE among compared conditioning methods.

Limitations and Future Directions

The current method focuses primarily on geometric outline deformation, not explicitly predicting typographic components like kerning pairs or hinting instructions. Future work is suggested to extend the framework to jointly predict glyph geometry, spacing behavior, and font-level layout tables. The training dataset could be expanded to include additional public or licensed variable font sources for increased stylistic diversity. Finally, the axis-conditioned deformation framework suggests potential extension beyond typography to structured vector graphics.

Key Technical Details:

(Input Representation)

The total sequence length is defined as N = Ng + 4, where Ng is the number of control points and 4 represents the phantom points. Coordinates are normalized to a common reference frame, and on-curve/off-curve indicators are encoded as scalar features.

(Multi-Axis Interaction)

The overall weight of a tuple at design-space configuration v is obtained by combining per-axis weights multiplicatively: λg,t(v) = ∑a=1 wt,a(va). The final glyph outline is given by Pg(v) = P(0)g + ∑t=1 λg,t(v) ∆g,t.

(Complexity)

The overall training complexity scales as O(NP2d), where P is the number of control points in a glyph and d is the latent dimensionality.

Here are the specific improvements to AI systems that can be derived from the NIV (Neural Axis Variations) method, along with a description of what those improved systems can achieve:

The NIV method provides a powerful framework for synthesizing continuous geometric variations directly into functional variable font files, moving beyond static image generation. The core innovation lies in predicting per-point displacement vectors conditioned on a multi-axis design space using an interaction-aware Property Embedding mechanism.

Here are the specific improvements and capabilities:

1. ---

2. Generating Fully Functional Variable Fonts from Static Inputs (The Core Improvement)

3. The system can automatically convert any static font file (e.g.,.ttf,.otf) into a fully functional OpenType variable font file supporting continuous interpolation along user-defined semantic axes (weight, width, slant, optical size).

4. This capability is achieved by training a sequence-to-sequence geometric model that predicts the necessary outline deformations (delta vectors on control points) conditioned on the desired axis settings.

5. The system operates directly on vector glyph geometry and outputs standard OpenType variable font files (.ttf), enabling immediate use in existing rendering engines and design software via continuous axis sliders, bypassing the labor-intensive manual specification of gvar glyph variation data.

6. ---

7. Modeling Higher-Order Cross-Axis Interactions (The Structural Improvement)

8. The system can accurately predict complex, non-separable geometric deformations by explicitly learning the joint, nonlinear processing of multiple design axes simultaneously via the novel Property Embedding mechanism and multi-head self-attention blocks.

9. This allows the AI to handle realistic typographic phenomena where variations are not independent (e.g., how increasing weight affects stroke thickness relative to width expansion), leading to geometrically coherent and stylistically consistent interpolations across a unified design space, as demonstrated by the superior performance in joint axis evaluation compared to additive per-axis decomposition baselines.

10. ---

11. Robust Generalization Across Unseen Inputs (The Versatility Improvement)

12. The system demonstrates strong generalization capabilities beyond its training distribution, including:

13. Synthesizing variations for high-complexity CJK glyphs (thousands of points).

14. Generating coherent deformations for stylistically distinct typefaces (e.g., Brush Script, UnifrakturMaguntia) and entirely out-of-distribution inputs like person handwriting samples, even when the input has no resemblance to training data.

15. Predicting axis variations for unseen Unicode code points using only geometric outline information, suggesting a deep understanding of fundamental typographic structure rather than memorized character shapes.

16. ---

17. Automated Synthesis for Diverse Graphic Domains (The Expansion Improvement)

18. The underlying neural deformation framework can be extended beyond typography to synthesize continuous parametric variations in other structured geometric domains (e.g., symbolic pictograms or SVG-like vector graphics).

19. This extension is possible because the model learns a unified deformation space, suggesting it can apply its learned principles of structure preservation and axis-driven adjustment to generate novel, stylistically coherent vector graphics based on static outlines.

20. ---

21. Reduced Engineering Overhead for Type Design (The Efficiency Improvement)

22. For designers or developers needing rapid style iteration, the system drastically reduces the manual engineering effort required to create variable fonts by automating the generation of gvar tables from a single static source, saving significant time and reducing errors associated with manually defining complex variation data.