Fine grained rewards improve Chinese text rendering in AI images
Decompose Radicals, Then Reward: Fine-Grained Inspection for Accurate Chinese Text Rendering
Computer Vision and Pattern Recognition
Summary
Generating clear and accurate Chinese text with AI images is hard because Chinese characters are made of smaller parts arranged in complex ways. Current methods treat each character as a whole and don’t check these parts carefully, so the AI might guess characters that look similar but are not correct inside. The authors created a new way to break down and check characters piece by piece using a detailed description system called IDS, which helps the AI get more precise feedback. This better feedback makes the AI create characters that not only look right but have the right internal structure.
What this means in practice
- •For graphic designers: Generate Chinese text images with accurate character structure for branding and publishing.
- •For machine learning engineers: Improve training of text-to-image models to produce visually and semantically correct Chinese characters.
Authors
Yazhen Xie, Xingsong Ye, Zhineng Chen
Abstract
Rendering accurate Chinese text remains challenging for text-to-image models. Existing OCR-based reinforcement-learning rewards compare decoded transcripts with target strings. Such rewards overlook the compositional nature of Chinese writing: an ideograph consists of reusable components arranged through explicit spatial relations, yet OCR evaluates it as an atomic character. Consequently, visually different radical-level errors may receive equally coarse feedback, encouraging glyphs that merely resemble the target instead of faithfully reproducing its internal structure. We employ Ideographic Description Sequences (IDS), which comprise spatial operators and character components, and train an expert IDS recognizer to transcribe rendered Chinese text into this representation. Building on this recognizer, we introduce IDSpect, which deterministically decomposes the target text into IDS tokens and aligns crop-level visual IDS predictions with the target sequence. Globally unique token credit makes this comparison robust to the order of detected text regions. Combined with a whole-character semantic reward, IDSpect supplies fine-grained credit with component and spatial-relation without changing the image generator or adding inference-time cost. Experiments with GRPO post-training of Qwen-Image demonstrate that IDSpect achieves leading structural quality and semantic alignment on LongText and GenTextEval.