Open-Form Geometric Style Transfer

Adar Elad; Gal Metzer; Omer Dahary; Min Lu; Daniel Cohen-Or

cs.CVOct 19, 2026

Open-Form Geometric Style Transfer

Authors: Adar Elad, Gal Metzer, Omer Dahary, Min Lu, Daniel Cohen-Or

Organizations: Tel Aviv University · Shenzhen University

Abstract

A geometric style learned from one family of shapes is typically applied to new shapes from the same family. We consider a more challenging setting: learning the geometric design language of a stylized Latin alphabet and applying it to new forms, from Greek glyphs to entirely different structures such as human poses. We introduce Open-Form Geometric Style Transfer, where the forms to which a learned geometric style can be applied are not restricted to the shape family in which that style was observed. This setting is deceptively challenging: the difficulty is not merely learning a style, but learning the style without learning the shape family in which it was observed. We represent target forms by structural sketches and disentangle structural adherence from style realization using a Geometry LoRA and a style-specific LoRA. Critically, open-form transfer requires the Geometry LoRA to be universal. To achieve this, we devise a geometry-enrichment strategy that constructs a large and purposefully diverse collection of synthetic structural sketches spanning a broad range of unfamiliar geometries. Our experiments and ablations demonstrate the importance of both disentanglement and this universal geometry prior, enabling geometric styles learned from glyphs, pictograms, and other shape families to be transferred to substantially different, unseen forms.

Figures & tables

Panel of Figure 1 — Figure 1: Open-Form Geometric Style Transfer. Given a small set of examples sharing a geometric style (left), we learn their common geometric design language and apply it to unseen structural sketches. The resulting forms (right) span challenging and unrelated shape families, illustrating open-form transfer beyond the family in which the style was originally observed.

Panel of Figure 2 — Figure 2: Open-Form Geometric Style Transfer. Given a set of glyph examples sharing a geometric style (left), we learn to apply that style to new structural sketches (right). The target forms are unseen during style learning and may belong to entirely different shape families.

Panel of Figure 3 — Figure 3: Method overview. (i) GEE expands a small vocabulary of familiar glyphs into a large corpus of unfamiliar geometries paired with style-suppressed structural sketches. (ii) The Universal Geometry LoRA (UGL) is trained once on this corpus and then frozen, providing a general geometry-following prior. (iii) Each Style LoRA is trained on top of the frozen UGL using only a few style examples, with geometry reinforcement preserving its open-form capability. (iv) At inference, UGL and the Style LoRA are applied jointly over the open-form inputs

Panel of Figure 4 — Figure 4: GEE overview. (a) Geometry generation recursively merges and removes glyph components in font $i$ , recording a font-independent recipe. (b) Reapplying the recipe to font $j$ produces the same underlying construction in another style. (c) Skeleton extraction yields a conditioning sketch from $j$ , which is paired with the target glyph from $i$ .

Panel of Figure 5 — Figure 5: Graph extraction. The steps of extracting a structural graph from a glyph, shown for two glyphs. Columns labelled as referenced in Sec. 3.2 .

Figure 6: Component removal. The steps of removing one structural component, shown for two glyphs. Columns labelled as referenced in Sec. 3.2 .

Figure 7: Merging. The steps of merging the same pair of glyphs in two styles. Columns labelled as referenced in Sec. 3.2 .

Panel of Figure 8 — Figure 8: Path to geometric generalization. Each pair shows an unseen sketch and the output in different styles. w/o UGL trains the Style LoRA without UGL, and w/o GEE trains the UGL without GEE. In both, the failure patterns of the geometry are marked. Ours follows the same sketches while preserving each style.

Panel of Figure 9 — Figure 9: In-domain comparisons. Each row shows a sketch held out from that style’s training examples, two reference examples of the style, the outputs of the baselines, the ground truth, and our output. The font styles (top) are compared with the dedicated font-generation models, the colorful styles (bottom) with the general-purpose image models.

Panel of Figure 10 — Figure 10: Out-of-domain results. Each result is shown directly below its framed structural sketch. Each sketch lies outside the shape family used to learn the corresponding style. Top row: a single sculptural style applied to glyphs from unseen scripts, a bridge, and free-form shapes. Reference examples are included to illustrate the diversity within the learned style. Bottom row: three additional styles: two pictogram styles (left & right) and a second sculptural style (middle)

Panel of Figure 11 — Figure 11: Near-domain results. Each row shows a non-Latin glyph sketch, two reference examples of the style, the outputs of the baselines, and our output next to an enlarged view of the region marked in red. The font styles (top) are compared with the dedicated font-generation models, the colorful styles (bottom) with the general-purpose image models.

Panel of Figure 13 — Figure 13: Further comparisons. Each row shows the structural sketch, three examples of the style (framed), the result of the compared method, and ours. Left: B-LoRA [ 10 ] transfers palette and texture but not the structure. Right: ZipLoRA [ 24 ] , applied to our UGL and a Style LoRA trained without it, follows the sketch but weakens the style.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 14 — Figure 14: Comparison with AiS. For each framed sketch, the result of AiS [ 19 ] (top) and ours (bottom), for two pictogram styles and a sculptural style.

Branch	% of samples	Geometry $c_{I}$	Style $c_{T}$
Fully cond.	70%	$c_{I}$	$c_{T}$
Geometry cond.	20%	$c_{I}$	$\varnothing_{T}$
Style cond.	5%	$\varnothing_{I}$	$c_{T}$
Uncond.	5%	$\varnothing_{I}$	$\varnothing_{T}$

Appendix

Table 2: Conditioning branches during training and their sampling frequencies.

Panel of Figure 15 — Figure 15: The effect of the two guidance scales of Eq. 1 ; $s_{T}$ grows left to right, $s_{I}$ top to bottom. The framed cell is our default operating point. The structural sketch and three reference examples of the style are shown below the grid.

	UGL (stage 1)	Style LoRA (stage 2)
LoRA rank / $\alpha$	$192$ / $192$	$32$ / $32$
Optimisation steps	$6{,}000$	$2{,}500$
Batch $\times$ grad. accum.	$16\times 4$	$4\times 2$
Warm-up steps	$400$	$50$
Learning rate	$10^{-4}$	$10^{-4}$
Held out	$5\%$ recipes	$6$ rows per style

Appendix

Table 3: Training settings of the two stages.

Panel of Figure 17 — Figure 17: Additional results. Each row shows three examples of a style (left) and our results for unseen structural sketches, each shown above its result. The rows cover font and pictogram styles.

Image Style Transfer

Image style transfer changes an image’s visual appearance to match an artistic, photographic, or other reference while preserving its content.

See all 61 papers

AnchorGen: Anchored Optimization for Customizable Generative 3D Design
Oct 5, 2026Hantao Zhang, Oliver Heinimann, Jieke Wu +3Constrained Generative Modeling 3D Shape Editing
CVLab, EPFL, Switzerland
VecGlypher: Unified Vector Glyph Generation with Language Models
Feb 25, 2026Xiaoke Huang, Bhavul Gauri, Kam Woh Ng +12Vector Graphics Generation Text-to-SVG Generation
Meta AI · UC, Santa Cruz
Internalizing Geometric Law: Learning from Solver Residuals for Precision-Critical Generation
Jun 8, 2026Rafael Cabral, Pang Zixi, Ziyi Shou +1Reward Shaping LLM-Based Program Synthesis
Huawei Celia Team

Oct 5, 2026J/K move · Enter open · S save

AnchorGen: Anchored Optimization for Customizable Generative 3D Design

Hantao Zhang, Oliver Heinimann, Jieke Wu, Yingxuan You+2

CVLab, EPFL, Switzerland

Constrained Generative Modeling 3D Shape Editing 3D Shape Generation Flow-Based Generative Modeling

Open paper AskPDF arXiv

Engineering design often starts from a 2D sketch that fixes style and proportions, yet the subsequent 3D shape optimization relies on learned generative priors to keep the geometry valid. However, these priors are agnostic to the sketch: while they admit a valid design by correcting a drifted proposal back to its training distribution, they often correct it towards the high-density region, ignoring the specified design. We introduce AnchorGen, a rectified-flow framework trained unconditionally on the concatenated shape and sketch latents of paired data. The learned manifold represents the joint distribution of shape-sketch pairs, so constraining the sketch component restricts the iterate to the sub-manifold of shapes consistent with a target style. Since training employs no conditioning signal, the constraint is imposed at inference: gradient descent optimizes the shape latent to minimize a differentiable drag surrogate, while constraining the sketch latent to remain close to the target sketch via a token-wise cosine penalty. A single model thereby supports design-preserving optimization, dimensionally explicit design edits, and sketch-only synthesis.

Abstract

Image Style Transfer

Explore similar work

AnchorGen: Anchored Optimization for Customizable Generative 3D Design

VecGlypher: Unified Vector Glyph Generation with Language Models

Internalizing Geometric Law: Learning from Solver Residuals for Precision-Critical Generation