NIV: Neural Axis Variations auto-converts static fonts to variable fonts

NIV (Neural Axis Variations) is a new neural approach from researchers Nadav Benedek, Ariel Shamir, and Ohad Fried that eliminates the labor-intensive process of manually creating variable fonts from static ones. Traditional variable font design requires expert typographers to specify glyph variation data for each semantic axis (weight, width, slant, optical size). NIV automates this by operating directly on vector glyph geometry and using a Property Embedding mechanism that captures interactions between multiple design axes, enabling consistent multi-axis variation in a single framework.

The model was trained on a newly constructed dataset derived from Google Fonts, comprising over one million variation tuples. It generalizes well beyond its training data: it handles unseen code points, unseen font styles, complex CJK (Chinese-Japanese-Korean) glyphs, and even out-of-distribution handwriting inputs. The output is a standard variable font file that can be used with any existing rendering engine supporting continuous interpolation. The team has released the dataset, training and inference code, and trained models on GitHub. Beyond typography, the approach demonstrates how structured geometric objects with parametric variation can be synthesized using neural deformations, opening possibilities for other design domains.