Neural Image Abstraction Using Long Smoothing B-Splines Journal: TOG Volume: 44 6 225 12 DOI: 10.1145/3763345 CCS: Computing methodologies Non-photorealistic rendering CCS: Computing methodologies Rasterization CCS: Computing methodologies Parametric curve and surface models CCS: Computing methodolo

Daniel Berio, Michael Stroh, Sylvain Calinon, Frederic Fol Leymarie, Oliver Deussen, Ariel Shamir · 2025

Read Neural Image Abstraction Using Long Smoothing B-Splines Journal: TOG Volume: 44 6 225 12 DOI: 10.1145/3763345 CCS: Computing methodologies Non-photorealistic rendering CCS: Computing methodologies Rasterization CCS: Computing methodologies Parametric curve and surface models CCS: Computing methodolo by Daniel Berio, Michael Stroh, Sylvain Calinon, Frederic Fol Leymarie, Oliver Deussen, Ariel Shamir (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-05360.txt.

Introduction

Daniel Berio and co-authors fold long smoothing B-splines into a differentiable vector graphics pipeline, so learned image systems can draw smooth paths of any length. Vector art built from short segments tends to look jagged unless a smoothing step is added.

Their linear mapping keeps the whole chain trainable by gradient methods. Derivative-based costs let users trade fidelity for simplicity, and steer style in both shape and image terms. The team shows four uses, from art generation to vectorization work, and the pipeline fits recent tools in both fields well.

Catalog record

CREATORDaniel Berio, Michael Stroh, Sylvain Calinon, Frederic Fol Leymarie, Oliver Deussen, Ariel Shamir
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-05360

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Neural Image Abstraction Using Long Smoothing B-Splines Journal: TOG Volume: 44 6 225 12 DOI: 10.1145/3763345 CCS: Computing methodologies Non-photorealistic rendering CCS: Computing methodologies Rasterization CCS: Computing methodologies Parametric curve and surface models CCS: Computing methodologies Neural networks CCS: Applied computing Fine arts Daniel Berio, Michael Stroh, Sylvain Calinon, Frederic Fol Leymarie, Oliver Deussen, Ariel Shamir Abstract . We integrate smoothing B-splines into a standard differentiable vector graphics (DiffVG) pipeline through linear mapping, and show how this can be used to generate smooth and arbitrarily long paths within image-based deep learning systems. We take advantage of derivative-based smoothing costs for parametric control of fidelity vs. simplicity tradeoffs, while also enabling stylization control in geometric and image spaces. The proposed pipeline is compatible with recent vector graphics generation and vectorization methods. We demonstrate the versatility of our approach with four applications aimed at the generation of stylized vector graphics: stylized space-filling path generation, stroke-based image abstraction, closed-area image abstraction, and stylized text generation. Figure 1. Our method allows to optimize long and smooth B-Spline curves using DiffVG rasterization pipelines. Applications range from text stylization to image abstraction and vectorization. † † cc-license: by 1. Introduction T he ability to produce long, smooth curves is central to a variety of design and artistic tasks. These include freehand drawing, sketching, calligraphy, typography, logo design as well as image abstractions into compositions of organic, flowing, or blobby shapes. Our aim is to enable the generation of vector graphic outputs that allow these types of designs, while taking advantage of recent advances in gradient-based image generation, stylization, and understanding. Developments in differentiable vector graphics (DiffVG) rasterization have enabled gradient-based optimization methods that leverage complex image-space losses to drive image generation, stylization and abstraction methods. Most existing approaches rely on the method of Li et al. 2020 , which implements differentiable rasterization for a large subset of elements of the Scalable Vector Graphics (SVG) standard, including piecewise cubic and quadratic Bézier curves. Most of these methods directly optimize Bézier curves, but even with additional smoothing penalties they do not provide guarantees of continuity across segments, which limits their ability to represent long, smooth and expressive strokes. Our work is based on two observations. First, alternative spline parametrizations such as B-spline ( De Boor 2001 ) or Catmull-Rom ( DeRose and Barsky 1988 ) provide inherent continuity constraints in their definition. Second, the conversion of such curves to Bézier curves is a linear transformation, making their integration into existing DiffVG pipelines a matter of an additional matrix multiplication. Although these curve parameterizations are well established, to the best of our knowledge, their integration into DiffVG remains largely unexplored. In our work we focus on uniform B-splines for their simplicity, high-order continuity and analytic properties ( Farin 2001 ) . This enables a straightforward implementation of derivative-based smoothing criteria that are well known in the fairing and motor-control/robotics domains, but most importantly support our goals of generating long stylized curves within a DiffVG pipeline. We define B-splines with control-polygons consisting of series of “key-points” and convert these to piecewise cubic Bézier curves for rendering. Since this transformation is linear and rendering is differentiable, gradients from image-space losses can be back-propagated to the key-points. We treat stroke width as a third curve dimension, where each curve control point can be assigned an independent stroke radius, enabling smooth variations similar to that seen in physical brush strokes ( Fujioka and Kano 2007 ) . Allowing the stroke width to vanish also presents an effective way to alter the number of visible strokes required for an image abstraction. Our method operates with both open and closed curves, supporting the generation of closed and organic-looking areas. We present four different applications for our method: abstract space filling curves (Section 4.1 ), sketch-based stylization (Section 4.2 ), abstract image vectorization with color quantization (Section 4.3 ) as well as text stylization and calligram generation with a novel legibility cost (Section 4.4 ). We provide a practical implementation of smoothing B-splines that can be directly integrated into DiffVG pipelines and demonstrate how this enables long and expressive curves while maintaining flexible geometric and stylistic control. Working code and examples for our method are available at github.com/colormotor/calligraph . 2. Related work 2.1. Smooth and stylized curve generation Long and stylized strokes have been explored used in the literature for applications including image stylization ( Kaplan and Bosch 2005 ; Wong and Takahashi 2011 ; Tong et al. 2025 ) , text-based stylization ( Maharik et al. 2011 ) , and fabrication ( Liu et al. 2017 ; Yang et al. 2021 ) . To widen and enhance such applications, our method also enables the generation of long and smooth strokes through the use of neural-driven image-based costs. Smoothing is achieved by minimizing the squared magnitude of higher-order positional derivatives. In the motor control literature, it is well established that the kinematics of hand and arm movements can be modeled by optimizing performance criteria ( Flash and Hogan 1998 ) . The so-called minimum square derivative models have been successfully applied to handwriting and curved motion by minimizing third-order derivatives (jerk) ( Flash and Hogan 1985 ) and fourth-order derivatives (snap) ( Edelman and Flash 1987 ) . Similar minimum principles are widely employed for smooth motion control in drones ( Mellinger and Kumar 2011 ; Ren and Kry 2019 ) and robots ( Toussaint 2017 ; Todorov 2004 ) , as well as in statistics for smoothing noisy data ( Reinsch 1967 ; Eilers and Marx 1996 ) . Similar principles of smooth motion and continuity have also been used for curve fairing, where a “fair” curve is typically one that exhibits a smooth variation of the curvature ( Farin 2001 ) . In this context, jerk has been adopted as an approximation for curvature variation ( Lu 2015 ; Pottmann 1990 ; Meier and Nowacki 1987 ) , while snap serves as an approximation for transverse distributed load ( Meier and Nowacki 1987 ) . In an extensive body of work, Egerstedt and Martin 2009 develop “dynamic splines” that formulate polynomial splines through optimal control of linear systems. Berio et al. 2017 use similar principles for the interactive generation of stylized paths similar to the ones seen in graffiti art and calligraphy with applications similar to ours. Kano et al. 2003 study the relations between dynamic splines and B-splines and in a collection of work, they develop an optimal formulation of B-splines ( Kano et al. 2005 ) applied to generate motion paths and curves similar to those found in Japanese calligraphy ( Matsukida and Fujioka 2013 ; Fujioka et al. 2006 ) . Our approach is strongly inspired by the B-spline construction initially proposed by Kano et al. 2005 , but we extend their formulation to support DiffVG and demonstrate its flexibility for generative and stylization settings. 2.2. DiffVG and applications In recent years, differentiable rendering has enabled the use of large pretrained vision and generative imaging models with 3D ( Kato et al. 2020 ; Tewari et al. 2020 ; Worchel and Alexa 2023 ) and 2D ( Mihai and Hare 2021 ; Li et al. 2020 ; Worchel and Alexa 2023

Read the complete text