Research map / Image generation / Text-to-image diffusion
GlyphAnchor: Enhancing Visual Text Rendering via Position-Anchored Glyph Priors
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 GlyphAnchor, which adds rendered glyph patch images as extra condition tokens whose RoPE position IDs are anchored to layout boxes in diffusion transformers, trained with staged SFT and text-aware DiffusionNFT post-training, and introduces InfoTextBench.
Problem
Diffusion image generation and editing models render long, complex, densely arranged text and rare characters inaccurately; prior glyph-prior methods are insufficiently accurate and robust.
Approach
Glyph patch images are encoded as extra condition tokens and assigned position IDs from the target layout via the backbone's native RoPE, leaving the attention pattern unchanged; training shifts from ground-truth crops to rendered glyphs, followed by NFT post-training with text-layout, style, and fusion rewards.
Key results
- On ChineseWord, FireRed-Image-Edit-1.1 + GlyphAnchor raises total accuracy from 0.565 to 0.801 and Qwen-Image-Edit-2511 + GlyphAnchor from 0.496 to 0.807.
- On InfoTextBench, Z-Image + GlyphAnchor improves phrase hit from 0.807 to 0.885 and overall visual quality from 7.092 to 8.822.
- On LongTextBench, FireRed-Image-Edit-1.1 + GlyphAnchor improves average from 0.897 to 0.985.
Surprising findings
- Glyph patch SFT without positional anchoring gives large gains, but adding RoPE anchoring further improves all InfoTextBench metrics, with the largest gain in phrase hit rate.
- Full-canvas glyph references weaken prompt following and suppress decorative detail and background texture.
Limitations
- Requires a layout input, obtained from user, templates, or an MLLM planner, adding a dependency on layout quality.
- Overall visual quality and reward signals rely on MLLM judges (Qwen3.5-122B-A10B), and InfoTextBench is small (133 samples).
Experimental setup
Datasets: web-curated text-rich images template-constructed text images Table of General Standard Chinese Characters
Benchmarks: LongTextBench OneIGBench-text CVTG-2K ChineseWord InfoTextBench
Metrics: word accuracy NED CLIP score WAC word recall word precision word F1 phrase hit rate overall visual quality
Baselines: Qwen-Image Qwen-Image-2512 Qwen-Image-Edit-2511 FireRed-Image-Edit-1.1 Z-Image GLM-Image LongCat-Image Seedream 4.5 Nano Banana Pro TextCrafter TextDiffuser-2 AnyText
Builds on: DiffusionNFT RoPE FlashAttention Qwen-Image
Closest papers by meaning
- GlyphMastero: A Glyph Encoder for High-Fidelity Scene Text Editing (CVPR 2025 · similarity 0.89)
- SteerVTE: Seamless Video Text Editing with Style and Glyph Control (NeurIPS 2026 · similarity 0.87)
- TextCtrl: Diffusion-based Scene Text Editing with Prior Guidance Control (NeurIPS 2024 · similarity 0.87)
- GlyphDraw2: Automatic Generation of Complex Glyph Posters with Diffusion Models and Large Language Models (AAAI 2025 · similarity 0.86)
- VecGlypher: Unified Vector Glyph Generation with Language Models (CVPR 2026 · similarity 0.86)
- GlyphSR: A Simple Glyph-Aware Framework for Scene Text Image Super-Resolution (AAAI 2025 · similarity 0.86)
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.