Decompose Radicals, Then Reward: Fine-Grained Inspection for Accurate Chinese Text Rendering
Yazhen Xie, Xingsong Ye, Zhineng Chen
Fudan University · Shanghai Key Laboratory of Multimodal Embodied AI
Published September 29, 2026 · on the feed October 1, 2026 · ▲ 5 · arXiv 2609.37569
The 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.
The Brief
2:08A short podcast summary of the paper. Written and voiced by AI.
Transcript
Decompose Radicals, Then Reward. Fine-Grained Inspection for Accurate Chinese Text Rendering. By Yazhen Xie and two others, from Fudan University and the Shanghai Key Laboratory of Multimodal Embodied AI. Published on the twenty-ninth of September. No code repository is linked to it.
Text-to-image models still struggle to render Chinese text accurately. The reinforcement-learning rewards used to fix this are usually OCR-based: decode the rendered text and compare the transcript with the target string. That misses how Chinese writing is built. An ideograph is made of reusable components, arranged through explicit spatial relations, yet OCR treats each character as one indivisible unit. Errors at the level of a radical can look quite different from each other and still receive the same coarse feedback. The result is a reward that encourages glyphs which merely resemble the target, rather than ones that reproduce its internal structure faithfully.
The fix is to grade at the level of components. The authors use Ideographic Description Sequences, or IDS, which write a character as spatial operators and character components. They train an expert IDS recognizer that transcribes rendered Chinese text into that representation, so the reward can see the parts and how they are arranged.
On top of that recognizer sits IDSpect. It deterministically decomposes the target text into IDS tokens, then aligns the visual IDS predictions made on each crop with that target sequence. Credit is assigned to globally unique tokens, which makes the comparison robust to the order in which text regions are detected. IDSpect is combined with a semantic reward on whole characters, and together they supply fine-grained credit for components and for spatial relations. None of this changes the image generator, and none of it adds cost at inference time.
Experiments use GRPO post-training of Qwen-Image. Evaluated on LongText and GenTextEval, IDSpect achieves leading structural quality and leading semantic alignment.
Two characters with quite different mistakes inside them can earn the same coarse score from an OCR check. Breaking the target into radicals and their arrangement gives the reward a way to tell those apart, and the image model something specific to learn from.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a reward that sees a Chinese character as a single, indivisible block.
You are post-training an image generator to write Chinese text inside pictures. The reward reads the picture with OCR and compares the transcript with the text you asked for.
The road ahead
- 1 · The problem ~6 min An OCR reward reads each Chinese character as one whole label, so it cannot give credit for getting part of a character right.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Score each rendered character by its parts and their layout, read straight from the image and matched with the target, alongside a whole-character reward.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Train a parts recognizer, decompose the target, align crop predictions with it, add a semantic reward, and post-train the generator with GRPO.
- 6 · The experiment ~6 min You’ll guess: After GRPO post-training of Qwen-Image, how did IDSpect do on structural quality and semantic alignment on LongText and GenTextEval?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.