Source-linked AI summary
Decompose Radicals, Then Reward: Fine-Grained Inspection for Accurate Chinese Text Rendering
Yazhen Xie, Xingsong Ye, Zhineng Chen
TL;DR
OCR-based rewards treat Chinese ideographs as atomic transcripts, limiting feedback on their internal structure. IDSpect combines semantic text fidelity with aligned IDS-token matching, and GRPO post-training with IDSpect improves structural quality and semantic alignment on LongText-Bench and GenTextEval-Bench.
Problem
OCR-based rewards evaluate Chinese ideographs as atomic transcripts, giving indistinguishable feedback to structurally distinct errors despite their component-and-relation composition.
Source: §1 · pp. 1–2Method
IDSpect combines OCR transcript fidelity with an expert recognizer’s crop-level IDS predictions aligned to deterministic target IDS tokens for fine-grained structural credit.
Source: §1 · p. 2, §3.1 · p. 3, §3.3 · p. 5Results
Takeaways & Limitations
Abstract
from arXiv · showhide
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.
1 Introduction
Chinese ideographs combine reusable components through spatial relations, but OCR rewards evaluate them as atomic transcripts, obscuring partial structural correctness. IDSpect adds target-conditioned IDS matching to GRPO rewards and improves structural quality and semantic alignment over OCR rewards and TextPecker on LongText and GenTextEval.
1 Introduction
- OCR rewards can give unchanged feedback when one radical is repaired, because transcript evaluation treats ideographs as atomic labels.Visually different internal errors can therefore receive indistinguishable feedback.Source: §1 · pp. 1–2
- For the same malformed glyph, TextPecker marks structural anomaly but cannot distinguish its correct from incorrect internal structure, while IDSpect credits matched IDS tokens.The example assigns IDSpect a standalone IDS reward of 0.22.Source: §1 · p. 2, Figure 1 · p. 2
- IDSpect trains an expert IDS recognizer and aligns visual IDS predictions with deterministic target IDS tokens, combining structural credit with character-level semantic fidelity.Its IDS represents spatial operators and character components, allowing partial credit when the whole character is incorrect.Source: §1 · pp. 2–3
- On LongText and GenTextEval, IDSpect improves structural quality and semantic alignment over general OCR-based rewards and TextPecker.The contribution statement reports both quality dimensions across both benchmarks.Source: §1 · p. 3
2 Related Work
Related work covers visual-text generation and post-training alongside structure-aware Chinese recognition, motivating the use of decomposition as a generative reward rather than solely as a recognition aid.
2.1 Visual text generation and post-training
- Specialized encoders and glyph conditioning improve visual text rendering, while TextPecker’s “#” treatment of incorrect characters remains coarse [19] [22].This motivates measuring edit progress within a target ideograph rather than assigning only a whole-glyph or region-level quality judgment.Source: §2.1 · p. 3
2.2 Structure-aware Chinese text recognition
- Radicals, strokes, and IDS have supported recognition of rare or unseen Chinese characters [33]; this work instead uses an IDS reader as a reward model for generation.Reward utility also depends on whether intermediate scores usefully rank partially correct and malformed glyphs.Source: §2.2 · p. 3
3 Method
The method pairs whole-character semantic feedback with IDS-based structural feedback to train Chinese text rendering. It directly recognizes IDS from rendered text, aligns crop predictions to target tokens, and uses uniquely credited matches for a compositional reward.
3.1 Problem formulation
- The semantic reward measures normalized character edit similarity between the target and OCR transcript, but treats each character atomically and offers limited feedback on partial structure.A substring match receives full credit even if the OCR output contains additional characters.Source: §3.1 · p. 3
3.2 Balanced IDS recognizer
- The IDS recognizer directly maps rendered text images to sequences of spatial operators and character components, rather than decomposing a recognized character transcript.Its decoder vocabulary comes from recursive IDS decompositions of Chinese characters covered by GB 18030–2022, with decomposition depth capped at 10.Source: §3.2 · p. 4
- Balanced synthetic training data enrich rare ideographs and discourage character-frequency or linguistic shortcuts, promoting component-grounded IDS predictions.IDSynth-1M pairs rendered strings with IDS sequences using a rare-ideograph-enriched corpus with balanced frequencies across covered characters.Source: §3.2 · p. 5
3.3 Target-conditioned compositional reward
- Crop-level IDS predictions are aligned to candidate target spans with a global assignment, making matching crop-order independent while preserving token order within each crop.Each target position can receive exact-match credit only once; component and operator matches can improve the reward even when atomic OCR output is unchanged.Source: §3.3 · p. 5
- The compositional token-level F1 reward is combined with semantic reward for GRPO post-training, and both reward branches are discarded afterward.IDSpect therefore adds no inference-time cost to the generator.Source: §3.3 · p. 5
4 Experiments
Experiments post-train Qwen-Image with GRPO and compare IDSpect against OCR- and TextPecker-guided rewards on LongText and GenTextEval. IDSpect improves structural and semantic metrics, while ablations show that target-based, crop-wise IDS alignment matters for semantic alignment.
4.1 Experimental setup
- The experiments post-train Qwen-Image with Flow-GRPO, using LoRA adapters and equal weights for the semantic and IDS rewards.The IDS recognizer is trained on IDSynth-1M, with approximately uniform character sampling to reduce lexical co-occurrence and encourage visual parsing.Source: §4.1 · p. 6
4.2 Experimental Results
- On LongText, IDSpect achieves the highest Qua. and Sem. scores, 0.975 and 0.928, while its Avg. score of 0.972 is comparable to TextPecker; on GenTextEval, its 0.979 Qua. and 0.911 Sem. exceed TextPecker by 0.006 and 0.014.On GenTextEval, the gains over OCR reward are 0.026 Qua. and 0.037 Sem.Source: §4.2 · p. 6
- In qualitative comparisons, IDSpect renders requested main and secondary text more completely, preserves character structure, and maintains the layout of multiple text regions.The advantage is especially apparent in longer lines on signs and posters, where structural errors can accumulate.Source: §4.2 · pp. 6–7, Figure 3 · p. 7
4.3 Ablation of IDS reward construction
- Per-crop consistency scores 0.980 Qua. but 0.901 Sem., below target-based crop-wise alignment’s 0.911 Sem.The OCR-derived IDS reference measures structural consistency with the OCR transcript rather than directly enforcing agreement with the requested text.Source: §4.3 · p. 7
- On GenTextEval, crop-wise target alignment reaches 0.911 Sem., exceeding direct concatenation by 0.025 and character-wise matching by 0.012, with Qua. of 0.979.Crop-wise alignment associates detected crops with target spans and gives each target token unique credit while preserving within-crop IDS token order.Source: §4.3 · p. 8, Table 2 · p. 8
5 Conclusion
The conclusion presents IDSpect as a target-conditioned compositional reward that combines semantic fidelity with component- and spatial-operator-level supervision. Qwen-Image experiments show improved structural quality and semantic alignment on two benchmarks, while broader character and script coverage remains future work.
5 Conclusion
- IDSpect combines character-level semantic fidelity with fine-grained supervision over ideograph components and spatial operators.The conclusion frames decomposing ideographs into reusable components and explicit spatial relations as a practical route to more faithful Chinese text generation.Source: §5 · p. 8
- With GRPO post-training of Qwen-Image, IDSpect improves structural quality and semantic alignment on LongText-Bench and GenTextEval-Bench over general OCR-based rewards.The conclusion also reports stronger supervision than glyph-level anomaly-aware evaluation.Source: §5 · p. 8
- The stated future direction is broader Chinese character coverage and extension to multilingual scripts such as Japanese and Korean.Source: §5 · p. 8