# Visual Text Rendering

The authors built VTR-Bench, a benchmark that embeds prescribed text and its carriers into 300 prompts (1,202 text blocks) across advertising, science, user interfaces, culture, and daily life, scores text fidelity by carrier-specific transcription and word error rate (WER) and scene and motion fulfillment by a 20-question prompt-specific chain of query, and adds a Keyframe-Guided Agentic Framework in which a Director agent coordinates image generation, motion planning, and visual feedback; across 11 state-of-the-art models they find widespread difficulty in rendering scene text, with the best model reaching an overall WER of 0.250, while the framework cuts overall WER on Minimax H3 from 0.4468 to 0.3015 (a 32.5% relative reduction) and raises Video Score.

The authors built VTR-Bench, a benchmark that embeds prescribed text and its carriers into 300 prompts (1,202 text blocks) across advertising, science, user interfaces, culture, and daily life, scores text fidelity by carrier-specific transcription and word error rate (WER) and scene and motion fulfillment by a 20-question prompt-specific chain of query, and adds a Keyframe-Guided Agentic Framework in which a Director agent coordinates image generation, motion planning, and visual feedback; across 11 state-of-the-art models they find widespread difficulty in rendering scene text, with the best model reaching an overall WER of 0.250, while the framework cuts overall WER on Minimax H3 from 0.4468 to 0.3015 (a 32.5% relative reduction) and raises Video Score.

The work introduces IDSpect: it recursively decomposes target Chinese characters into Ideographic Description Sequence (IDS) tokens using the Unicode 16.0 BabelStone lexicon, trains an expert IDS recognizer built on an SVTRv2 backbone with an NRTR-style decoder to transcribe rendered text regions directly into IDS sequences, and fuses a token-level F1 with globally unique token credit against a whole-character semantic reward, so that GRPO post-training of Qwen-Image improves structural quality and semantic alignment on LongText and GenTextEval without changing the image generator or adding inference-time cost.

The work introduces IDSpect: it recursively decomposes target Chinese characters into Ideographic Description Sequence (IDS) tokens using the Unicode 16.0 BabelStone lexicon, trains an expert IDS recognizer built on an SVTRv2 backbone with an NRTR-style decoder to transcribe rendered text regions directly into IDS sequences, and fuses a token-level F1 with globally unique token credit against a whole-character semantic reward, so that GRPO post-training of Qwen-Image improves structural quality and semantic alignment on LongText and GenTextEval without changing the image generator or adding inference-time cost.