title: >- [Paper Note] LoGAN: Multilingual Font Localization with Generative Agents description: >- [ECCV 2026][Image Generation][Font Localization] This paper presents LoGAN, the first agentic framework coordinated by a Vision-Language Model (VLM) for end-to-end multilingual font and logo localization. By decoupling generation into glyph synthesis, kerning transfer, style fine-tuning, and texture expansion, it supports high-fidelity vectorized rendering across 27 languages with a 52.9% Production-Ready rate on real-world movie logos. tags: - "ECCV 2026" - "Image Generation" - "Font Localization" - "VLM Agent" - "Glyph Diffusion Model" - "Typography & Kerning" - "Multilingual Generation" date: 2026-09-19 content_hash: 47b73b234af024bc

LoGAN: Multilingual Font Localization with Generative Agents¶

Conference: ECCV 2026
Paper: ECCV Official
Full Text Cache: /Users/zy/workspace/paper_cache/ECCV2026/eccv-4037.txt
Area: Image Generation
Keywords: Font Localization, VLM Agent, Glyph Diffusion Model, Kerning Transfer, Multilingual Rendering

TL;DR¶

LoGAN introduces a VLM-agent-coordinated framework that decomposes multilingual font localization into modular stages—glyph generation, kerning and layout transfer, style fine-tuning, and texture expansion—achieving high-fidelity vector font rendering across 27 languages (including CJK) and outperforming existing text-editing and diffusion baselines.

Background & Motivation¶

Fonts and stylized letterforms serve as foundational visual anchors in graphic design, branding, and entertainment marketing. However, the vast majority of highly stylized artistic fonts support only a limited set of languages—predominantly offering standard Latin character sets. When creative assets such as movie title logos need to be adapted for global audiences, creative teams face a severe "font localization bottleneck." Artistic fonts frequently employ idiosyncratic geometric motifs, custom stroke weights, dramatic terminals, intricate surface textures, and handcrafted spatial dynamics (kerning). Transferring these unconstrained design languages across drastically different scripts—such as mapping simple Latin strokes to structurally dense CJK ideograms without sacrificing legibility—remains an enormous typographic challenge. In a sample study of 400 popular movie titles, 44% of logos were created directly using fonts; when no matching localized font exists, graphic designers are forced either to compromise stylistic coherence by using generic substitute fonts or to laboriously synthesize each character manually.

Recent breakthroughs in generative diffusion models (e.g., FLUX, Nano-Banana, GPT-Image) suggest the potential for automated text rendering and image-guided editing. Nonetheless, applying these models directly to multilingual font localization exposes three fundamental limitations: First, specialized font generation models (such as FontDiffuser or VecGlypher) operate almost exclusively at the isolated single-glyph or monochrome SVG level, completely lacking the ability to infer inter-character spacing, holistic layout, or rich textural material. Second, general-purpose image editing and text-to-image models attempt to synthesize content, layout, font style, and texture in a single unconstrained forward pass. This entanglement frequently results in layout instability, unintended character rescaling, inconsistent word arrangements, and style drift. Third, current generative pipelines exhibit poor multilingual coverage; models trained predominantly on Latin text suffer from severe structural breakdowns, missing strokes, and illegible hallucinations when prompted with complex CJK ideograms or non-Latin scripts.

The central insight of this paper is that end-to-end one-pass generation is inherently mismatched with the structured, multi-factor nature of typography. Core idea: decompose multilingual font localization into a three-stage agentic pipeline—design extraction, design transfer, and design generation—coordinated by a central Vision-Language Model (VLM) agent that orchestrates a 2B MMDiT cross-lingual glyph diffusion model, a geometric contour-based kerning transfer algorithm, dynamic style LoRA adaptation, and an inpainting-based texture expansion model.

Method¶

Overall Architecture¶

LoGAN takes as input a source reference font or logo image, a target text translation, and optional user instructions; it outputs a fully rendered, stylized text image alongside an editable SVG vector asset. The entire workflow is governed by a central VLM agent (implemented by default using GPT-5.2, with verified equivalence using open-source Qwen3-VL-32B-Instruct) acting across three distinct roles: Interpreter (decomposing user prompts and parsing reference visual attributes), Planner (transferring layout metadata and deciding tool execution), and Judge (evaluating character correctness and typographic fidelity with automated retries). The execution pipeline spans three consecutive stages: Design Extraction, Design Transfer, and Design Generation.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Reference Image + Target Translation"] --> B["Stage 1: AutoCrop Character-Level Detection & Kerning Extraction<br/>Connected components + VLM grouping + watertight bounding boxes"]
    B --> C["Stage 2: VLM Agent Design Transfer Planning<br/>Character-attribute mapping + style LoRA trigger + texture branch"]
    C --> D["Stage 3.1: Cross-Lingual Glyph Generation & Style Adaptation<br/>2B MMDiT multimodal attention + optional reference LoRA"]
    D --> E["Stage 3.2: Glyph Arrangement & Heuristic Kerning Transfer<br/>Four abstract contour shapes + line break and alignment"]
    E --> F["Stage 3.3: Texture Expansion & Vectorization Synthesis<br/>FLUX-Fill texture recovery + Potrace/VectorMagic SVG vectorization"]
    F --> G{"VLM Agent Multimodal Evaluation & Verification<br/>OCR inspection / missing stroke check / color consistency"}
    G -->|Failure Detected / Retry| D
    G -->|Pass Verification| H["Output: Localized Rendered Image & Vectorized SVG Font"]

Key Designs¶

1. AutoCrop Character-Level Cropping and Attribute Parsing: Resolving Multi-Component and Diacritic Segmentation

To feed clean, isolated character exemplars into the downstream glyph diffusion backbone and accurately record inter-glyph physical spacing, reference fonts must be cleanly cropped into individual glyphs. Traditional text detectors (such as DBNet or PaddleOCR) typically output word- or line-level bounding boxes, while character detectors like CRAFT fail on artistic fonts featuring fragmented strokes or domain shifts. LoGAN designs a training-free multilingual character cropping algorithm termed AutoCrop: - It first extracts all connected components and their binary masks via standard topological component labeling; - It renders numeric index tags over each extracted component; - It invokes a VLM with OCR context to group related component indices by character identity, clustering disconnected radicals in CJK characters or detached accents in Latin scripts (e.g., À, ü) into unified, watertight bounding boxes; - Finally, it sorts bounding boxes in standard reading order (top-to-bottom, left-to-right) and records the exact horizontal pixel gap between adjacent boxes into a kerning lookup dictionary.

2. Cross-Lingual Multimodal Glyph Diffusion (MMDiT) with Balanced Sampling: Overcoming Long-Tailed Script Imbalance

To overcome the fragility of prior font generators on complex non-Latin scripts, LoGAN develops a specialized 2B-parameter glyph diffusion backbone built on the Multimodal Diffusion Transformer (MMDiT, SD3.5-medium architecture). Inside each transformer block, the joint attention mechanism is augmented with a third modality dedicated to reference glyph conditioning: - Reference glyphs are encoded using SigLIP-2 (so400m-patch16-256), where embeddings of $k$ reference glyphs are concatenated and projected through a Perceiver Sampler to produce a fixed-length conditioning token sequence of shape $(N, 256, 1152)$; - Glyph prompt encoding relies on ByT5-Large to process raw byte-level character identities without BPE subword fragmentation; - Latent image representations are mapped using AuraDiffusion's 16-channel VAE. Because raw typography collections suffer from an extreme long-tail bias toward Latin scripts, the authors introduce a Language-Aware Char Sampling strategy: training batches balance same-language character pairs (50%) with forced cross-language pairs (50%, e.g., Latin to CJK or Greek to Cyrillic). Additionally, case-sensitive inversion and diacritic-aware normalization ensure the model learns style-invariant typographic representations (stroke weight, terminal geometry, slant, and contrast) decoupled from specific glyph topologies.

3. Abstract Geometric Contours and Heuristic Kerning Transfer: Typographic Spacing Beyond Direct Lookup

Kerning refers to the fine-grained horizontal space adjustment between adjacent letter pairs to achieve uniform perceived visual density (e.g., "WA" requires much tighter spacing than "HO"). Prior generative models either ignore spacing or apply uniform gaps, yielding unbalanced text blocks. LoGAN introduces a geometry-guided kerning inference mechanism: - The left and right contours of any character across languages are classified into four abstract shapes: Straight ($|$), Circular ($\circ$), Ascending Diagonal ($/$), and Descending Diagonal ($\backslash$). All compact CJK ideograms are mapped to Straight ($|$), while characters like "O" exhibit circular contours and "A" exhibits diagonal contours; - Based on expert typographic studies, the model defines a strict visual compactness ordering across contour pairings: $$S(|, |) > S(\circ, |) = S(|, \circ) > S(\circ, \circ) > S(/, \backslash) \ge S(\backslash, /)$$ where $S(\cdot)$ represents the target pixel spacing between the right contour of the preceding character and the left contour of the succeeding character. When a target letter pair exists in the reference image, the cached pixel measurement from AutoCrop is directly reused; for unseen character pairs, the spacing is inferred using the geometric ordering scaled by measured reference baselines, bounded by a global minimum spacing threshold.

4. Reference Style LoRA Fine-Tuning and Inpainting-Based Texture Expansion: High-Fidelity Style and Material Recovery

For unique artistic movie logos, zero-shot generation can introduce subtle stroke discrepancies. LoGAN introduces two modular enhancement mechanisms: - Style-Finetuning LoRA: When high-fidelity preservation is demanded, AutoCrop crops the reference glyphs into 512×512 square inputs to fine-tune a lightweight LoRA (Rank=64, Alpha=64, ~159.5M trainable parameters) on the 2B diffusion backbone. Running for only 800 steps (15–30 minutes on a single A100 GPU), this one-time tuning locks in distinct terminal cuts and stylistic accents; - Texture-Expansion Model: To preserve complex surface textures (e.g., gold foil, weathered stone, metallic rust), LoGAN fine-tunes a LoRA (~44.8M parameters) on FLUX.1-Fill-dev using 542 PBR materials. Given a textured text image, the model uncrops and inverts the text mask to synthesize a seamless, full-resolution background texture map, which is subsequently composited back onto the vectorized SVG glyph paths generated via Potrace or VectorMagic.

A Worked Example¶

Consider an input reference logo for the movie title "CARAMELO" (a bold, bright-yellow sans-serif font with tight letter spacing) to be localized into Chinese "可拉米洛". 1. Extraction: AutoCrop identifies 8 Latin letters, extracts the hex color #f1c40f, identifies Straight/Circular boundary contours, records measured gaps (e.g., "C-A" at 8px, "A-R" at 8px, "R-A" at 3px), and flags has_texture: false. 2. Transfer: The VLM agent identifies the Chinese target characters as four Straight-boundary ($|$) ideograms, maintains #f1c40f fill, and calculates a balanced uniform kerning gap of 10px. 3. Generation: The 2B MMDiT model synthesizes the four individual Chinese characters at 512×512 resolution; the arrangement module spaces them by 10px, applies #f1c40f, and executes Potrace vectorization; the VLM Judge verifies stroke integrity and releases the final localized SVG logo.

Loss & Training¶

The 2B glyph generation backbone is trained under a rectified flow objective matching the velocity field: $$\mathcal{L}_{\text{RF}} = \mathbb{E}_{t, x_0, x_1} \left[ \| v_\theta(x_t, t, c_{\text{text}}, c_{\text{ref}}) - (x_1 - x_0) \|^2 \right]$$ where $x_0 \sim \mathcal{N}(0, I)$, $x_1$ represents the target glyph latent, $x_t = (1-t)x_0 + tx_1$, $c_{\text{text}}$ is the ByT5 character embedding, and $c_{\text{ref}}$ is the Perceiver Sampler reference glyph embedding. - Pre-training: Distributed across 8 A100 (80GB) GPUs using FSDP in bfloat16. Pre-training begins at 256×256 resolution for 450k iterations (batch size 1024, AdamW learning rate $1.0 \times 10^{-4}$ with 1,000 warmup steps), followed by 512×512 resolution for another 450k iterations (batch size 640). Total pre-training takes approximately 15 days. - Texture Expansion Training: Fine-tuned on 2 A100 (40GB) GPUs using FLUX.1-Fill-dev at 1536×1536 resolution for 20,000 steps (batch size 2, learning rate $1.0 \times 10^{-4}$), taking approximately 12 hours.

Key Experimental Results¶

Main Results¶

Quantitative evaluations are conducted across two benchmarks: a synthetic movie title dataset (300 pairs across 50 unseen fonts, covering Latin, Greek, Cyrillic, and CJK) and a real-world movie logo dataset (700 monochrome logos). Baselines include text-specialized diffusion models (AnyText2, TextDiffuser-2, Flux-Text) and state-of-the-art general image editing models (Flux-Kontext Pro, Nano-Banana). Metrics comprise pixel-level reconstruction (PSNR, SSIM) and deep perceptual alignment (LPIPS, DreamSim, CLIP Score).

Table 1: Quantitative comparison on multilingual localized text rendering (Table 2 in paper)

Language Split (#)	Model	Editing Paradigm	PSNR ↑	SSIM ↑	LPIPS ↓	DreamSim ↓	CLIP Score ↑
CJK Split (42)	AnyText2	Text-specialized	6.894	0.329	0.825	0.510	0.811
	TextDiffuser-2	Text-specialized	6.460	0.449	0.882	0.604	0.666
	Flux-Text	Text-specialized	6.900	0.580	0.706	0.374	0.858
	Flux-Kontext Pro	General Editing	7.050	0.635	0.667	0.162	0.801
	Nano-Banana	General Editing	7.658	0.622	0.449	0.084	0.932
	LoGAN w/o LoRA (GPT-5.2)	Agentic Framework	9.052	0.657	0.588	0.101	0.920
	LoGAN (Qwen3-VL-32B)	Agentic Framework	8.186	0.671	0.536	0.081	0.929
	LoGAN (GPT-5.2 Full)	Agentic Framework	9.494	0.692	0.446	0.076	0.948
Non-CJK Split (254)	AnyText2	Text-specialized	7.123	0.458	0.806	0.524	0.812
	TextDiffuser-2	Text-specialized	6.875	0.541	0.862	0.631	0.658
	Flux-Text	Text-specialized	6.973	0.560	0.706	0.444	0.838
	Flux-Kontext Pro	General Editing	7.784	0.596	0.607	0.158	0.876
	Nano-Banana	General Editing	8.988	0.638	0.481	0.084	0.922
	LoGAN w/o LoRA (GPT-5.2)	Agentic Framework	9.094	0.654	0.444	0.085	0.932
	LoGAN (Qwen3-VL-32B)	Agentic Framework	9.162	0.664	0.451	0.109	0.927
	LoGAN (GPT-5.2 Full)	Agentic Framework	10.077	0.684	0.373	0.072	0.944

Ablation Study & Expert Evaluation¶

To isolate layout and spacing variations from raw glyph generation quality, the authors cropped localized outputs into individual glyphs and evaluated them directly against Nano-Banana (Table 3 in paper). In addition, performance was dissected across textured and untextured splits (Table 4 in paper). Finally, a professional font designer conducted a blind review of 700 localized logos for "Production-Ready" commercial usability.

Table 2: Single-glyph analysis and texture expansion ablations (Tables 3 & 4 in paper)

Evaluation Scope	Experimental Setting	Method	PSNR ↑	SSIM ↑	LPIPS ↓	DreamSim ↓	CLIP ↑
Single-Glyph Level (Table 3)	Tight crops excluding kerning	Nano-Banana	15.75	0.726	0.179	0.122	0.958
	Tight crops excluding kerning	LoGAN (Ours)	18.67	0.802	0.121	0.094	0.968
Texture Split (Table 4)	With Texture (50 pairs)	Nano-Banana	10.65	0.606	0.483	0.100	0.916
	With Texture (50 pairs)	LoGAN (Ours)	11.88	0.659	0.378	0.118	0.934
	Without Texture (250 pairs)	Nano-Banana	7.85	0.647	0.504	0.081	0.924
	Without Texture (250 pairs)	LoGAN (Ours)	9.00	0.699	0.391	0.049	0.951

Key Findings¶

Severe Degradation of Baseline Models on CJK: Existing text models suffer catastrophic performance drops on CJK characters (AnyText2 SSIM drops to 0.329), frequently outputting distorted or hallucinated strokes. In contrast, LoGAN achieves 9.494 PSNR and 0.692 SSIM on CJK, matching its non-CJK performance due to balanced cross-lingual pre-training.
Style LoRA Bridges Fine-Grained Domain Gaps: Even in zero-shot mode without LoRA, LoGAN outperforms all baselines (e.g., CJK DreamSim of 0.101 vs. Nano-Banana's 0.084). Injecting the lightweight style LoRA further boosts PSNR from 9.052 to 9.494, confirming that reference-guided tuning eliminates subtle pixel shifts on artistic terminals.
Decisive Professional Production-Ready Assessment: In expert designer ratings (3-point scale), LoGAN scored 2.8 on style transfer, 2.2 on kerning/layout, and a perfect 3.0 on translation accuracy (vs. 1.0, 1.0, and 19/80 for Nano-Banana). On 700 real-world movie logos, 52.9% of LoGAN outputs were graded as Production-Ready, whereas none (0%) of Nano-Banana's outputs passed.
Pipeline Robustness: Under an extensive stress test across 1,800 real movie logo localization runs, LoGAN achieved a 97% success rate (only 3% failure), with failures primarily traced to non-compliant JSON formatting from upstream VLM calls.

Highlights & Insights¶

Decoupled Architecture Beats Monolithic Diffusion: Decoupling the entangled typography factors (glyph shape, spacing/kerning, color, and texture) into modular expert components supervised by a VLM agent fundamentally resolves the hallucinations and layout instability common to end-to-end models.
Heuristic Typographic Ordering Meets Deep Learning: Instead of relying on noisy regression for inter-character spacing, abstracting glyphs into four geometric boundary contours ($|,\ \circ,\ /,\ \backslash$) and applying typographic partial ordering yields robust, natural kerning with zero training overhead.
Byte-Level ByT5 Conditioning: Leveraging ByT5-Large instead of subword tokenizers avoids token fragmentation issues when handling rare ideograms and diacritical marks, ensuring precise cross-lingual character alignment.

Limitations & Future Work¶

Connected Cursive Scripts and Ligatures: The pipeline relies on single-character cropping and recombining, making it less suitable for continuous cursive handwriting, Arabic scripts, or intricate calligraphy where strokes overlap across characters.
Multi-Font Composition and Complex 3D Warping: Currently optimized for single-font logos; scenes combining multiple hierarchical font families or text warped along 3D perspective curves remain challenging for automated agent layout.
LoRA Fine-Tuning Latency: While optional, one-time style LoRA training takes 15–30 minutes, which introduces friction for interactive design workflows; future work could explore hypernetworks or test-time in-context adaptation.

vs. Specialized Font Generators (FontDiffuser / VecGlypher): Prior works synthesize isolated glyphs or black-and-white SVGs without multi-character spacing, layout reasoning, or rich texture handling; LoGAN delivers a complete, production-grade end-to-end typographic pipeline.
vs. Text-Centric Generative Models (AnyText2 / TextDiffuser-2): These models struggle heavily with non-Latin scripts (especially CJK) and cannot replicate intricate artistic logo styles; LoGAN expands coverage across 27 languages via balanced cross-lingual sampling.
vs. General Image Inpainting / Editing (Nano-Banana / FLUX-Fill): Unconstrained editing models alter original layouts, drift in style, and produce character hallucinations; LoGAN grounds text editing with structured typographic primitives and automated VLM verification.

Rating¶

Novelty: ⭐⭐⭐⭐ [Pioneering agentic framework for end-to-end multilingual font localization with geometric kerning transfer]
Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Spans 27 languages, 300 synthetic title pairs, 700 real movie logos, and rigorous professional designer blind evaluation]
Writing Quality: ⭐⭐⭐⭐⭐ [Clear structural formulation, rigorous typographic grounding, and consistent qualitative and quantitative evidence]
Value: ⭐⭐⭐⭐⭐ [Directly tackles a high-friction industry bottleneck in global media localization with production-ready vector outputs]