Research map / Image generation / Text-to-image diffusion
BizGen: Advancing Article-level Visual Text Rendering for Infographics Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BizGen, a layout-guided cross-attention SDXL-based model with a retrieval-augmented data engine (Infographics-650K) for generating article-level visual text in infographics and slides.
Problem
Text-to-image models fail at article-level visual text rendering and ultra-dense layouts (tens to hundreds of sub-regions) required for infographics and slides, due to long context and scarce high-quality business-content data.
Approach
Decompose a long article-level prompt into region-wise prompts by cropping latent tokens per layer with layout-guided cross attention, generate training data via layer-wise retrieval-augmented infographic synthesis, and refine each region at inference with a layout-conditional CFG that applies per-layer guidance scales.
Key results
- On BizEval, BizGen reaches 92.98% text spelling precision for infographics with up to 10 layers, versus 41.33% for FLUX and 16.42% for DALL-E 3 in the same bucket.
- BizGen achieves over 25% absolute gain over FLUX on infographics with more than 20 layers.
- In a ~10-person user study, BizGen was preferred over FLUX, SD3 Large and DALL-E 3 on typography quality in 84.3%, 83.4% and 57.9% of cases respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase training data from Infographics-65K to Infographics-650K (ablation) | BizGen · BizEval infographics, visual text spelling precision for ≤10 layers | visual text spelling precision: 86.47 -> 92.98 | Helped |
| Use layout-guided rather than region-wise cross attention (ablation) | BizGen · BizEval infographics, ≤10 layers | visual text spelling precision: 84.91 -> 86.47 | Helped |
| Use layout-guided rather than region-wise cross attention (ablation) | BizGen · BizEval infographics | GPT-4o aesthetics score: 6.82 -> 7.20 | Helped |
| Increase training resolution from 1120 × 448 to 2240 × 896 (ablation) | BizGen · BizEval infographics, ≤10 layers | visual text spelling precision: 44.68 -> 86.47 | Helped |
| Apply layout-conditional classifier-free guidance to layers (ablation) | BizGen · BizEval infographics | layer artifact reduction | Helped |
| Use BizGen instead of FLUX | BizEval infographics · BizEval infographics | GPT-4o prompt following score: 4.11 -> 8.09 | Helped |
| Use BizGen instead of FLUX | BizEval slides · BizEval slides | GPT-4o aesthetics score: 7.16 -> 7.64 | Helped |
| Use BizGen instead of FLUX | BizEval slides · BizEval slides | GPT-4o prompt following score: 6.32 -> 7.61 | Helped |
Surprising findings
- FLUX outperformed DALL-E 3 and SD3 Large on infographics, which the authors hypothesize means FLUX may have been trained on infographic data.
Limitations
- Ultra-dense layouts and long prompts are costly, and performance drops sharply as layer count grows (e.g., above 20 layers).
- Relies on GPT-4o as judge for most quality metrics, and on synthetic data generated with GPT-4o, LLaVA and SDXL-based layer generation.
Experimental setup
Datasets: Infographics-650K Slides-500K internal infographic dataset (5,000+) LayerDiffuse transparent layer database (1M+ layers)
Benchmarks: BizEval
Metrics: visual text spelling precision layer generation success rate (LGSR) GPT-4o aesthetics score GPT-4o prompt following score user study win rate style consistency
Baselines: DALL-E 3 SD3 Large FLUX.1-dev Glyph-SDXL-v2 Ideogram 2.0
Builds on: Glyph-ByT5 SDXL LayerDiffuse Set-of-Mark
Closest papers by meaning
- Multimodal DeepResearcher: Generating Text-Chart Interleaved Reports from Scratch with Agentic Framework (AAAI 2026 · similarity 0.86)
- InfoChartQA: A Benchmark for Multimodal Question Answering on Infographic Charts (NeurIPS 2025 · similarity 0.85)
- ChartGalaxy: A Dataset for Infographic Chart Understanding and Generation (ICLR 2026 · similarity 0.85)
- From Words to Structured Visuals: A Benchmark and Framework for Text-to-Diagram Generation and Editing (CVPR 2025 · similarity 0.85)
- VisInteract: Towards Dynamic Interactive Text-to-Visualization under Imperfect Queries (NeurIPS 2026 · similarity 0.84)
- DesignDiffusion: High-Quality Text-to-Design Image Generation with Diffusion Models (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.