LoGAN: Multilingual Font Localization with Generative Agents

Zhuoning Yuan; Ta-Ying Cheng; Benjamin Klein

cs.CVSep 7, 2026

LoGAN: Multilingual Font Localization with Generative Agents

Authors: Zhuoning Yuan, Ta-Ying Cheng, Benjamin Klein

Organizations: Netflix

Abstract

Localizing a font into new languages is a highly intricate task requiring precise design adaptation of glyphs, color/texture, and spacing/kerning, from source to target languages. Most existing methods focus on single glyph generation with limited capability in handling multilingual font rendering. In this work, we propose LoGAN, a VLM-based agentic framework for few-shot multilingual font localization, which takes in a small number of individual glyphs from a font or letters from a logo and uses them to generate complete character sets in other languages. LoGAN breaks down this task into multiple components: a glyph-level diffusion model, a style finetuning module, a spacing and kerning transfer algorithm, and a texture expansion model, with a VLM agent coordinator. LoGAN achieves broad language coverage for font localization with various styles, including Chinese/Korean/Japanese (CJK). We evaluate our approach on both font and real-world logo datasets spanning more than 27 languages and compare it against both specialized font generation and state-of-the-art image editing models with strong text rendering capabilities (e.g., FLUX, Nano-Banana). Our approach yields higher glyph fidelity while maintaining better style, texture, and kerning consistency according to both quantitative and qualitative evaluations.

Figures & tables

Figure 1: Overview. We present LoGAN, an agentic framework that localizes fonts into various languages. LoGAN is able to generalize to various font styles, spacing/kernings, colors, and languages spanning across Latin, Greek, Cyrillic and CJK.

Figure 2: Challenges of current image generation models. (1) Font-based models primarily focus on glyph generation and fails to render multiple characters with proper spacing/kerning and consistent color/textures. (2) End-to-end image editing models introduces randomness towards style changes and arrangements. (3) All current models lack a wide range of multilingual support.

Figure 3: Method Overview. LoGAN takes a reference image and a target translation generates translated font with three stages: (1) Design Extraction, (2) Design Transfer, and (3) Design Generation. Style LoRA is trained during (2) and then frozen for inference during (3).

Shape (Symbol)	Description
Straight ( $\mathbf{\|}$ )	Vertical or implied vertical boundaries that restrict close kerning (e.g., both sides of “H”, right side of “E”, left side of “B”, etc. All CJK letters are also assumed to fall within this shape).
Circular ( $\mathbf{\circ}$ )	Curved boundaries that naturally accommodate tighter kerning (e.g., both sides of “O” or the left side of “C”).
$\nearrow$ Diagonal ( $\mathbf{/}$ )	Ascending diagonal strokes from bottom-left to top-right (e.g., the left side of “A” or the right side of “V”. Right side of “T” also belongs to this shape as you can draw an implicit $\nearrow$ for tighter kerning).
$\searrow$ Diagonal ( $\mathbf{\backslash}$ )	Descending diagonal strokes from top-left to bottom-right (e.g., the right side of “A” or the left side of “V”). Certain lowercase letters, such as “e” or “w”, are also treated as diagonal because their open contours provide spatial gaps for tighter kerning.

Table 1: Abstract Shapes for Kerning Heuristics.

Figure 4: Illustration of character-level cropping for Chinese characters “Beverly Hills Cop".

Figure 5: Statistics of language distribution on our internal font dataset.

Table 7

Lang. (#)	Model	Editing Type	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DSim $\downarrow$	CLIP $\uparrow$
CJK (42)	AnyText2 [ 53 ]	Text	6.894	0.329	0.825	0.510	0.811
	TextDiffuser-2 [ 12 ]	Text	6.460	0.449	0.882	0.604	0.666
	Flux-Text [ 35 ]	Text	6.900	0.580	0.706	0.374	0.858
	Flux-Kontext Pro [ 32 ]	Image	7.050	0.635	0.667	0.162	0.801
	Nano-Banana [ 24 ]	Image	7.658	0.622	0.588	0.101	0.920
	LoGAN w/o LoRA (GPT-5.2)	Text	9.052	0.657	0.449	0.084	0.932

Table 2: Qualitative Comparison. We measure PSNR, SSIM, LPIPS [ 68 ] , DreamSim (DSim) [ 20 ] , and CLIP similarity score [ 47 ] . Results are shown for Non-CJK (254 pairs) and CJK (42 pairs). Editing type indicates whether the model is trained for text rerendering. Red/yellow denote 1st/2nd.

Figure 6: Qualitative Comparison. We present the results of 12 localized logos spanning across 4 movie titles. Flux-Kontext often ignores the text guidance and does not perform any edits, especially for CJK languages. Nano-Banana often misses the translations and slightly alters the style. Our results performs accurate style and layout transfer while simultaneously translating the titles correctly.

Figure 7: Glyph Quality Comparison Against FontDiffuser . FontDiffuser fails to match the original style/color of the reference font.

Figure 8: Texture Expansion Qualitative Comparison. We present the qualitative results of our texture expansion approach against Nano-Banana [ 24 ] .

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DreamSim $\downarrow$	CLIP $\uparrow$
NB	15.75	0.726	0.179	0.122	0.958
LoGAN	18.67	0.802	0.121	0.094	0.968

Table 3: Glyph-level comparison for LoGAN and Nano-Banana.

Group	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DSim $\downarrow$	CLIP $\uparrow$
With Texture	Nano-Banana	10.65	0.606	0.483	0.100	0.916
With Texture	LoGAN	11.88	0.659	0.378	0.118	0.934
Withou texture	Nano-Banana	7.85	0.647	0.504	0.081	0.924
Withou texture	LoGAN	9.00	0.699	0.391	0.049	0.951

Table 4: Quantitative comparisons for LoGAN and Nano-Banana in texture/no texture split.

Noise Scheduler
Parameter	Value
Shift	1.0 (3.0 for 512x512)
Timesteps	1000
Transformer (SD3Transformer2DModel)
Sample size	128
Patch size	2

Table 5: Configurations for Glyph-Generation model.

Figure 11: Comparison of PNG to SVG quality using different approaches . The leftmost column shows the rasterized image (PNG), while the right columns display results from Potrace, ImageTrace, StarVector (8B), and VectorMagic. Note that colors may have shifted slightly due to the screenshot and copy-paste from the source file.

Figure 12: Qualitative Results. We present 15 more generated results from LoGAN across 5 titles from from recently released movies.

Lang. (#)	Model	Editing Type	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DSim $\downarrow$	CLIP $\uparrow$
CJK (42)	AnyText2 [ 53 ]	Text	6.894	0.329	0.825	0.510	0.811
	TextDiffuser-2 [ 12 ]	Text	6.460	0.449	0.882	0.604	0.666
	Flux-Text [ 35 ]	Text	6.900	0.580	0.706	0.374	0.858
	Flux-Kontext Pro [ 32 ]	Image	7.050	0.635	0.667	0.162	0.801
	Nano-Banana [ 24 ]	Image	7.658	0.622	0.588	0.101	0.920
	Nano-Banana Pro [ 24 ]	Image	9.055	0.635	0.429	0.082	0.949

Table 6: Qualitative Comparison. We measure PSNR, SSIM, LPIPS [ 68 ] , DreamSim (DSim) [ 20 ] , and CLIP similarity score [ 47 ] . Results are shown for Non-CJK (254 pairs) and CJK (42 pairs). Editing type indicates whether the model is trained for text rerendering. Red/yellow denote 1st/2nd.

Figure 13: Texture Expansion Results. We present more results of texture expansion from our synthetic testing dataset. Our model is able to extract detailed structures even with very little hint from the given image (e.g., "Lion King" example where the strokes are extremely thin). Zoom in for details on the textures.

Panel of Figure 14 — Figure 14: Prompt for Design Extraction.

Panel of Figure 15 — Figure 15: Prompt for Design Transfer.

Panel of Figure 16 — Figure 16: Prompt for Auto Cropping.

SmartFont: Dynamic Condition Allocation for Few-Shot Font Generation
Jun 11, 2026Zian Yang, Zixin WangFew-Shot Learning Diffusion-Based Image Generation
Fudan University
LogoDiffuser: Training-Free Multilingual Logo Generation and Stylization via Letter-Aware Attention Control
Mar 10, 2026Mingyu Kang, Hyein Seo, Yuna Jeong +2Attention Control in Diffusion Models Multimodal Diffusion Models
Department of Artificial Intelligence, Hanyang University · Department of Computer Science, Hanyang University
VecFontLLM: Anchor-Guided Direct Synthesis of Chinese Vector Fonts
Jul 19, 2026Hao Yuan, Yuxuan Luo, Xing Chen +1Vector Graphics Generation
Fuzhou University, China · Peking University, China

Jun 11, 2026J/K move · Enter open · S save

SmartFont: Dynamic Condition Allocation for Few-Shot Font Generation

Zian Yang, Zixin Wang

Fudan University

Few-Shot Learning Diffusion-Based Image Generation Conditional Diffusion Models

Open paper AskPDF arXiv

Few-shot font generation simultaneously requires global structural completeness and fine-grained local style fidelity. Existing methods usually either rely on global content-style modeling, which is robust but imperfectly disentangled, or emphasize component/local modeling, which captures fine details but relies heavily on local priors and reference coverage. We argue that the key challenge is not merely to learn purer conditions, but to organize complementary yet biased global and local conditions through multi-level allocation during generation. To this end, we propose SmartFont, a diffusion-based few-shot font generation framework that combines global content-style generation with weakly supervised local corrective experts. The local branch performs semantic-spatial allocation by learning expert-wise local concepts and semantically meaningful spatial maps under weak component supervision, enabling fine-grained correction without requiring explicit component-conditioned inference. On top of this, a denoising-state condition allocation module adaptively weights global content, global style, and local corrective feature across timesteps and injection blocks. Extensive experiments show that SmartFont achieves better global-local balance, improves glyph quality and local detail fidelity.

Abstract

Explore similar work

SmartFont: Dynamic Condition Allocation for Few-Shot Font Generation

LogoDiffuser: Training-Free Multilingual Logo Generation and Stylization via Letter-Aware Attention Control

VecFontLLM: Anchor-Guided Direct Synthesis of Chinese Vector Fonts