arXiv:2605.19320v1 [cs.CV] 19 May 2026
TextAlign: Preference Alignment for Text Rendering with Hierarchical Rewards Mingxuan Cui1∗ Jingpu Yang2∗ Fengxian Ji1∗ Qian Jiang3 Zhecheng Shi4 Jiaming Wang3 Zirui Song1 Fajri Koto1 Xiuying Chen1† 1 Mohamed bin Zayed University of Artificial Intelligence 2 Beihang University 3 Northeastern University 4 The Hong Kong University of Science and Technology (Guangzhou)
Abstract Faithful text rendering remains a persistent weakness of large text-to-image generative models, as it requires both semantic instruction following and fine-grained glyph-level structure. Prior methods often improve this ability through architecturespecific modules or encoder modifications, which complicate deployment across foundation models. We study text rendering as a post-training preference-alignment problem and propose TextAlign, a non-invasive framework that keeps the generator architecture unchanged. The key component is a hierarchical vision-language model (VLM)-based reward that decomposes rendering errors into global, word, and glyph levels, then converts binary defect judgments into a scalar preference signal. The resulting signal supports both Group Relative Policy Optimization (GRPO) and Direct Preference Optimization (DPO). Experiments on FLUX.1-dev and Z-Image-Turbo show consistent gains in OCR-based text accuracy without degrading general generation quality. Compared with strong foundation and textrendering baselines, including SD3.5, Qwen-Image, AnyText, and TextDiffuser, these results indicate that reward design offers a scalable alternative to model redesign for improving text rendering.
1
Introduction
Large text-to-image generative models [5, 1, 39, 50] have made substantial progress in synthesizing high-fidelity images from open-ended natural language instructions. However, faithfully rendering text inside images remains a persistent failure mode [44]. Unlike generic object or style generation, visual text rendering requires the model to satisfy a discrete symbolic constraint while also producing continuous visual structure: the intended words must appear, characters must be ordered correctly, glyph shapes must be recognizable, and the text must be placed naturally within the surrounding scene. This makes text rendering a stringent test of both instruction following and fine-grained visual precision. As illustrated in Fig. 1, successful rendering must hold across varied carriers, layouts, styles, and string lengths, rather than only in simplified centered-text settings. Existing methods largely address this problem through architecture-level interventions. Some approaches replace or augment text and image encoders to strengthen character awareness [16, 53, 18, 6], while others introduce additional control modules or glyph-level conditions to guide the spatial layout of rendered text [2, 48, 34, 36]. These designs have improved visual text synthesis, but they also introduce a practical limitation: the solution is often tied to a particular model architecture, conditioning interface, or control representation. As a result, transferring the same recipe to a new foundation ∗ Equal contribution. † Corresponding author.
40th Conference on Neural Information Processing Systems (NeurIPS 2026).
A vibrant travel journal opens to a page filled with sketches of mountains and maps, with "Fuel your adventures with our travel essentials." boldly inked in deep navy script across the bottom margin.
A rustic wooden sign hangs above a bustling farmers' market, it reads "Fresh Flavors -MARKET- Real Ingredients · Real Taste".
A weathered wooden sign leans against a moss-covered stone wall in a misty forest clearing. It reads, "Be brave enough to follow your own path."
A vibrant sports stadium at sunset, with fans cheering and banners waving, features dynamic graffiti-style text reading "Love the game? Celebrate with our official sports gear."
A cartoonish exaggerated character in a comically oversized hoodie sits cross-legged on a cluttered couch, displaying "STOP LOOKING AT ME LIKE THAT.", suggesting a chaotic yet cozy home environment.
A steampunk brass plaque with intricate engravings reads: "Between the lines, the world is not as it seems, and the words hold the key to unlocking the truth."
Figure 1: Text rendering results. Representative 720 × 720 samples generated by our aligned models. TextAlign renders legible and well-formed visual text across diverse carriers, styles, layouts, and text lengths while preserving coherent image content.
model can require non-trivial engineering and may disturb the pretrained generative prior that gives modern models their broad visual competence. We take a different view: text rendering can be treated as a post-training preference-alignment problem. Rather than redesigning the generator, we ask whether a frozen foundation model can be aligned toward better typographic behavior through a reward signal that accurately reflects rendering failures. This perspective is attractive because it matches the deployment pattern of modern generative models: the base architecture is kept intact, while post-training adjusts model behavior toward a desired capability. It also makes the problem model-agnostic, enabling the same alignment framework to be applied to different image generators without introducing task-specific trainable modules. To this end, we propose TextAlign, a non-invasive preference-alignment framework for visual text rendering. The central technical challenge is reward design. Text rendering errors are inherently hierarchical: a generated image may contain no readable text at all, may render the wrong set of words, or may preserve the words while corrupting individual characters. A single OCR score or binary text-presence reward cannot reliably distinguish these failure modes. TextAlign therefore decomposes rendering quality into three levels: a global level that detects missing or malformed text, a word level that captures dropped, inserted, or substituted words, and a glyph level that identifies character-level insertions, deletions, and substitutions. Each level is judged by a vision-language model (VLM) through structured binary defect indicators, which are aggregated into a scalar reward for preference optimization. The resulting reward can be used by both Group Relative Policy Optimization (GRPO) [28, 41] and Direct Preference Optimization (DPO) [25, 35]. This compatibility allows TextAlign to serve as a general post-training interface rather than a method bound to a single optimizer. Importantly, because the reward is defined over generated images and target text, it does not require architectural modification, auxiliary glyph-control branches, or model-specific supervision signals. We evaluate TextAlign on FLUX.1-dev [1] and Z-Image-Turbo [50], and compare against strong foundation-model references including SD3.5 [5] and Qwen-Image [39], as well as representative text-rendering systems such as TextDiffuser [2] and AnyText [34]. Across OCR-based metrics, both DPO and GRPO consistently improve text rendering accuracy; on the stronger Z-Image-Turbo backbone, the GRPO-aligned model achieves the best normalized edit score, precision, recall, and F1-score among all compared methods. These gains are not obtained by sacrificing the original image 2
prior: CLIPScore, ImageReward, PickScore, and HPSv3 remain competitive after alignment, and a human study further favors the aligned models in both text fidelity and visual integration. We also analyze robustness across text length, spatial placement, and ten visual categories, where the aligned model maintains stable performance rather than overfitting to short, centered, or category-specific text. Taken together, these results support the central claim of this paper: reliable visual text rendering can be improved through carefully designed reward modeling and post-training alignment, without changing the architecture of large image generative models.
2
Related work
Diffusion models [30, 10, 46, 11, 45, 33, 14, 13, 49] have been widely adopted in text-to-image generation. They use a diffusion process to gradually add noise to the image for training and then reverse the process for generation. Subsequent advances in latent-space modeling [26, 9], transformer backbones [24], and rectified-flow or flow-matching formulations [15, 17] have further improved the scalability and sample quality of this paradigm. At present, open text-to-image systems such as Stable Diffusion 3 [5], FLUX [1], Qwen-Image [39], and Z-Image [50] represent the frontier of large-scale image generation. Text rendering in text-to-image diffusion models has been advanced along several complementary directions. Much prior work injects layout or glyph information through auxiliary control modules [19, 2, 48, 34, 3, 4, 52, 18, 37, 51, 54, 42, 29, 36, 20]. For example, TextDiffuser [2] first generates character-level segmentation masks and then performs mask-conditioned generation with characteraware supervision. Another line of work employs specialized text or image encoders [16, 53, 6]. For example, UDiffText [53] trains a character-aware text encoder with a codebook to replace the original text encoder. Yet another line explores complex training strategies [38, 32, 8]. For example, the AMO sampler [8, 31, 22] alternates between ODE overshooting and noise reintroduction to introduce Langevin dynamics correction, while adaptively controlling overshooting strength via cross-attention scores. Preference alignment in text-to-image diffusion models has been explored recently. One active thread adapts reinforcement learning with group-relative credit assignment under Group Relative Policy Optimization (GRPO) [28]; DanceGRPO [41] brings GRPO to visual generation and reports stable scaling across diffusion and rectified-flow models under diverse reward signals. A complementary thread avoids explicit reward modeling and instead aligns the generative policy with pairwise preferences: Diffusion-DPO [35] extends Direct Preference Optimization (DPO) [25, 47] to diffusion by defining a likelihood-compatible objective and shows strong improvements from human comparison data. However, these approaches typically optimize a single global reward or holistic preference signal, leaving fine-grained, attribute-level objectives such as accurate text rendering largely underexplored.
3
Method
3.1
Overview of TextAlign
We cast text rendering as a fine-grained alignment problem and propose TextAlign, a post-training preference-alignment framework dedicated to text rendering. In contrast to prior approaches that strengthen text rendering by replacing the text encoder or attaching auxiliary glyph-control modules, TextAlign introduces no additional trainable modules and leaves the network architecture of the foundation model unchanged; it sharpens typographic precision purely through preference alignment in the post-training stage. This non-invasive design offers two immediate benefits. First, the generality and diversity of the underlying image generator are preserved intact. Second, the same alignment recipe transfers across image foundation models of heterogeneous architectures without any modelspecific redesign. Formally, let πref denote the frozen reference (base) model and πθ the trainable policy, with conditioning prompts c ∼ D and samples x ∼ πθ (· | c). Given a reward function R(·) tailored to text rendering (Sec. 3.2), the TextAlign objective is max Ec∼D, x∼πθ (·|c) R(x, c) − β DKL πθ ∥ πref , θ
3
(1)
where β controls the deviation from the reference distribution. Eq. (1) is naturally compatible with mainstream preference-alignment paradigms: under GRPO [28, 41], the scalar reward is normalized within each sample group to obtain relative advantages; under the DPO paradigm [25, 35], the same reward is used to rank samples generated under the same condition and construct winner/loser preference pairs, which are then optimized with the Diffusion-DPO objective for diffusion-based image generators. We instantiate TextAlign with the same recipe on SD3, FLUX, Qwen-Image, and Z-Image to validate the architecture-agnostic nature of the framework. 3.2
TextAlign Reward Design
The crux of Eq. (1) lies in constructing a reward function R that faithfully reflects rendering quality. Failure modes of text rendering are inherently multi-level: from the most macroscopic case in which the image contains no readable text at all, through mid-scale errors such as dropped, inserted or substituted whole words, down to the most microscopic single-character substitutions, insertions and deletions. A reward that focuses on any single level inevitably leaves blind spots—supervising solely with OCR edit distance overlooks global structural failures, while a binary text-presence judgement cannot expose character-level defects. Motivated by this, we propose a three-level reward decomposition that covers, in a top-down fashion, the global, word and glyph granularities, explicitly classifying rendering failures and dispatching each level to an independent judgement by a vision-language model (VLM). For every generated sample together with its reference text y, the three levels are realized as three independent VLM calls, each focusing exclusively on its own failure modes and returning a set of binary indicators. The three sets of indicators are finally compressed into a single scalar by a unified aggregation function (Sec. 3.2.4) and consumed by Eq. (1). Fig. 2 provides an overview. 3.2.1
Global Level
The global level addresses the most fundamental questions: whether any readable text appears in the image and whether the overall glyph forms hold up. At this level the VLM independently produces two binary indicators: bno text records whether the image contains no recognizable text at all, and bmisshape records whether there exist characters of the correct identity but with severely distorted contours, including typical deformations such as broken strokes, imbalanced proportions, and warped baselines. The global level decides whether rendered content is present at all and therefore serves as a precondition for the two finer-grained levels that follow. 3.2.2
Word Level
The word level characterizes the model’s faithfulness to semantic units. We explicitly decompose word-level failures into three atomic indicators: bdrop word flags the presence of missing words, badd word the presence of extraneous words, and breplace word the presence of whole-word substitutions in which an intended word is rendered as a different word—typically manifested as two or more character errors occurring simultaneously. To produce these indicators, the VLM first extracts the rendered text from the image, normalizes case, removes punctuation, and performs a token-level alignment against the reference text y, yielding three mutually independent 0/1 signals. 3.2.3
Glyph Level
The glyph level targets the finest granularity of typographic precision. We symmetrically decompose character-level failures into three indicators: bdrop glyph flags missing characters, badd glyph flags inserted characters, and breplace glyph flags single-character substitutions in which the word length is preserved. This decomposition explicitly disentangles “length changes” from “content substitutions”: bdrop glyph and badd glyph require the surviving characters to remain a subsequence of the reference word, whereas breplace glyph requires the lengths to match with exactly one differing character, thereby avoiding any conflation with the whole-word replacement detected at the word level. 3.2.4
Reward Aggregation
The three levels jointly produce N =8 binary indicators {bk }N k=1 , each level emitted by an independent VLM call that is constrained to return a fixed-field structured response, thereby discretizing the 4
Generated Image
GLOBAL LEVEL (image-wide)
WORD LEVEL (semantic units)
VLM Detect
VLM Detect
misshape
no text
GLYPH LEVEL (character) VLM Detect
drop word
drop glyph
add word
add glyph
replace word
Reference text: "Welcome to NeurIPS"
no text example (to be filled)
replace glyph
Reference: Welcome to NeurIPS Rendered: Wlecome to at NeurIPS
misshape example (to be filled)
replace word (whole word)
Precondition
drop word (missing)
Reference: Welcome to NeurIPS Rendered: Welcom too NeurITS
add word (extra)
drop glyph (missing 'e')
add glyph (extra 'o')
replace glyph ('P' 'T')
→
Reward aggregation (8 binary indicators, 0-100 scale)
→
∑
parse fail sample discarded (excluded from GRPO group / DPO pair)
GRPO: group-normalized advantage Normalized Advantages
K Sample
X1 X2 X3
Xk
Low
Sample A (winner)
Sample B (loser)
Chosen
Sample B Xb
⋮
(z-score)
Xa
⋮
Xk
⋮
⋮ ⋮
X1 X2 X3
Probability Sample A
⋮
Normalize
DPO: winner/loser pairs within group Score
High
Reject
Figure 2: Our hierarchical reward mechanism. Given a generated image x and reference text y, three independent VLM calls produce binary indicators at the global, word and glyph levels, which are aggregated into a scalar reward R that drives either GRPO or DPO. model’s qualitative judgement into parsable signals. Let Nv ≤ N denote the number of indicators successfully parsed for a given sample. We define the scalar reward as: R =
Nv −
PNv
k=1 bk
Nv
,
(2)
which corresponds to the fraction of indicators reporting no defect. The aggregated scalar reward is used differently by the two optimizers: GRPO normalizes it within each group to obtain relative advantages, whereas DPO uses it only to construct winner/loser preference pairs under the same prompt. Thus, the global, word, and glyph levels require no optimizer-specific redesign; the same reward function supports both relative policy optimization and pairwise preference optimization. Whenever the structured response of any level fails to parse, we discard the entire sample and exclude it from both the group normalization in GRPO and the pair construction in DPO, in order to prevent noisy feedback from contaminating the gradients. Eq. (2) is simple in form, treats all defect categories on equal footing, and is robust to variations in the VLM output format. We adopt Qwen3.5-9B as the default reward VLM, and in Sec. 4 we further verify that the framework retains consistent performance when the reward VLM is replaced by models from the Qwen3-VL family.
4
Experiments
4.1
Experimental Setup
Dataset construction. Existing text rendering benchmarks are limited: prompts are typically short and templated, and target texts skew toward short, centered strings with uncontrolled semantic positions. This underrepresents real deployment difficulty, where text can be long, scenes visually complex, and target strings arbitrarily positioned. We therefore construct a more discriminative benchmark varying along three axes: text length, prompt complexity, and textual position. Samples further span ten visual-text carriers differing in layout complexity, text scale, background clutter, and typographic style. Full details are in Appendix A. Base models and baselines. We conduct the main preference-alignment experiments on two opensource foundation models, FLUX.1-dev [1] and Z-Image [50], under both DPO and GRPO. We additionally include SD3.5 [5] and Qwen-Image [39] as strong foundation-model references, and 5
Table 1: Text rendering accuracy under DPO and GRPO fine-tuning. We report OCR-based text accuracy metrics, including the normalized edit score, word-level precision, recall, f1-score, and exact word-set accuracy. Denotes the highest value in each column. model
NED
precision
recall
f1-score
accuracy
AnyText TextDiffuser SD3.5 Qwen-Image FLUX FLUX(Our DPO) FLUX(Our GRPO) Z-Image Z-Image(Our DPO) Z-Image(Our GRPO)
0.2857 0.3766 0.5876 0.8615 0.5768 0.5925 0.6075 0.8739 0.8830 0.8893
0.0014 0.5770 0.5242 0.8563 0.5746 0.5872 0.6070 0.8860 0.8994 0.9105
0.0008 0.3068 0.5146 0.8352 0.5301 0.5336 0.5758 0.8726 0.8790 0.8924
0.0008 0.3633 0.5045 0.8318 0.5315 0.5396 0.5829 0.8657 0.8762 0.8876
0.0000 0.0852 0.1019 0.4722 0.1667 0.1778 0.2037 0.5278 0.5352 0.5648
compare with two representative architecture-level text rendering methods: TextDiffuser [2], which guides generation via character-level segmentation masks, and AnyText [34], which injects glyph and positional information through ControlNet. Evaluation metrics. Our evaluation metrics are fully decoupled from the reward signals defined in Sec. 3.2. For textual accuracy, we extract the rendered text from generated images using PaddleOCRVL-1.5 [23] and report the normalized edit score (NED; higher is better) to measure character-level similarity to the target text, together with precision, recall, F1, and accuracy computed under the wordlevel matching protocol of TextDiffuser [2] to assess word-set-level agreement. For general generation quality, we report five external metrics, CLIPScore [7], Aesthetic Predictor [27], ImageReward [40], PickScore [12], and HPSv3 [21], to verify that the alignment process does not degrade the base model’s general synthesis capability. 4.2
Main Results
Table 1 reports OCR-based text rendering accuracy after DPO and GRPO fine-tuning. The evaluation is decoupled from the VLM reward used during training, covering the normalized edit score as well as word-level precision, recall, F1-score, and accuracy. On FLUX, both DPO and GRPO improve over the base model across all metrics, with GRPO yielding the larger gains. On the stronger Z-Image backbone, DPO also consistently improves the base model, and GRPO further achieves the best scores in all five OCR-based metrics, including exact word-set accuracy. These results suggest that the hierarchical reward is effective not only for relative policy optimization, but also for pairwise preference optimization. Compared with architecture-level text rendering methods such as AnyText and TextDiffuser, TextAlign achieves stronger text accuracy without introducing additional modules or modifying the foundation model architecture. 4.3
Generalization Analysis
To evaluate the generalization of TextAlign across diverse visual text scenarios, we report the categorywise performance of Z-Image (Our GRPO) over the ten representative categories described above. As shown in Table 3, the model maintains strong text rendering accuracy across categories, with an average NED of 0.8848 and an average F1-score of 0.8874. The strongest results are observed on basic, handwriting, and poster, while academic and scene show relatively lower exact-match accuracy, likely due to more complex layouts, smaller text regions, and stronger background interference. Meanwhile, CLIPScore, Aesthetics, and HPSv3 remain stable across categories, suggesting that TextAlign improves text readability without noticeably weakening category-specific style or visual semantics. We further analyze robustness to text length and spatial placement in Appendix B, where the aligned model maintains stable performance across different string lengths and layout positions. Additional qualitative examples in Appendix C show that the model can naturally embed legible text into diverse carriers such as posters, logos, stickers, handwriting, and natural scenes. 6
Table 2: General generation quality under DPO and GRPO fine-tuning. We report CLIPScore, ImageReward, PickScore, and HPSv3 to evaluate whether preference alignment preserves general image generation quality beyond text rendering accuracy. 1st 2nd 3rd. model
CLIPScore
ImageReward
PickScore
HPSv3
AnyText TextDiffuser SD3.5 Qwen-Image FLUX FLUX(Our DPO) FLUX(Our GRPO)
20.77 28.53 31.04 31.21 29.38 29.49 29.65
-0.902 -0.328 0.682 0.915 0.744 0.721 0.724
19.11 19.93 21.12 21.54 21.41 12.33 21.41
6.010 8.097 10.60 11.65 11.84 11.76 11.88
Z-Image
31.36
0.926
21.54
12.16
Z-Image(Our DPO)
31.56
0.898
21.54
12.12
Z-Image(Our GRPO)
31.47
0.928
21.55
12.13
Table 2 further evaluates whether these accuracy gains preserve general generation quality. Overall, the aligned models remain competitive in CLIPScore, aesthetic quality, ImageReward, and HPSv3, and the GRPO variants maintain strong PickScore values; the main exception is FLUX (Our DPO), whose PickScore decreases despite improved text accuracy. The qualitative results in Fig. 4 show a consistent trend, where aligned models generate more complete, clearer, and more readable text while preserving the scene, style, and layout specified by the prompt. We further report a user study in Fig. 3, following the twoquestion protocol of TextDiffuser [2] to evaluate text fidelity and visual integration. The aligned GRPO variants receive more votes than their corresponding base models on both criteria, with Z-Image (Our GRPO) obtaining the strongest overall preference, indicating that the OCR-based gains are also reflected in human perceptual judgments. 4.4
Figure 3: User study. Human preference votes on text fidelity and visual integration. Our GRPO-aligned models outperform prior baselines and base generators on both criteria, with Z-Image (Our GRPO) preferred most.
Evaluation on External Dataset
To test whether the gains from TextAlign transfer beyond our constructed benchmark, we further evaluate the same models on a 500-sample split of the external MARIO-Eval benchmark from TextDiffuser [2]. This evaluation uses the same OCR-based protocol as above, so the reward model used during alignment is not reused as the evaluator. As shown in Table 4, GRPO alignment improves both evaluated backbones on all reported text rendering metrics. For FLUX, the F1-score increases from 0.2437 to 0.2795 and the word-level exact-match accuracy increases from 0.1000 to 0.1240. For Z-Image, the aligned model also improves NED, precision, recall, F1-score, and accuracy over its base model. These results indicate that the hierarchical reward does not merely fit our in-domain evaluation set, but provides a transferable text-rendering signal that remains beneficial under an external prompt distribution. 4.5 Ablation Study Table 5 ablates the hierarchical reward by removing one level at a time from the full TextAlign reward. On FLUX.1-dev, the full reward achieves the best NED, precision, recall, and F1-score, and removing the glyph level leads to the largest drop in NED, indicating the importance of fine-grained character-level feedback. On Z-Image, the full reward gives the strongest precision, recall, and 7
Figure 4: Qualitative comparison of text rendering results. Given the same prompts, GRPOaligned FLUX and Z-Image produce more faithful and legible visual text while preserving the surrounding visual context.
F1-score, although some ablated variants slightly improve a single metric such as NED or strict accuracy. Overall, the three reward levels are complementary: global feedback stabilizes readable text structure, word-level feedback preserves semantic units, and glyph-level feedback refines character accuracy. 4.6
Qualitative Results and User Study
Figure 4 provides qualitative comparisons under the same prompts. Compared with the base models, the GRPO-aligned FLUX.1-dev and Z-Image-Turbo models generate more complete, clearer, and more readable text while preserving the surrounding scene, style, and layout specified by the prompt. These examples show that the alignment process improves text fidelity without visibly disrupting image composition. Additional category-wise examples are provided in Appendix C. We further conduct a user study following the two-question protocol of TextDiffuser [2], evaluating both text fidelity and visual integration. As shown in Fig. 3, the aligned GRPO variants receive more human preference votes than their corresponding base models on both criteria, with Z-Image-Turbo with GRPO alignment obtaining the strongest overall preference. This confirms that the improvements measured by OCR-based metrics are also reflected in human perceptual judgments.
5
Conclusion
We presented TextAlign, a non-invasive preference-alignment framework for improving visual text rendering in large text-to-image generative models. Rather than modifying the generator architecture or adding glyph-control branches, TextAlign keeps the foundation model intact and builds a reward 8
Table 3: Per-category performance of Z-Image (Our GRPO). We report text rendering accuracy and generation quality metrics across ten visual scene categories. 1st 2nd 3rd. Category
NED
precision
recall
f1-score
accuracy
CLIPScore
HPSv3
Poster Advertisement Cover
0.9099 0.8757 0.8712
0.9281 0.8963 0.9051
0.9033 0.8676 0.8924
0.9012 0.8651 0.8796
0.6481 0.5370 0.5370
32.14 31.84 30.00
12.61 11.71 12.63
Logo
0.9208
0.8960
0.8761
0.8762
0.6481
32.11
11.87
Sticker
0.8636
0.9359
0.8936
0.8937
0.5741
32.48
12.58
Handwriting
0.9205
0.9422
0.9485
0.9420
0.6852
29.89
11.89
Scene
0.8882
0.8886
0.8344
0.8518
0.4630
31.16
12.99
Basic
0.9091
0.9302
0.9617
0.9340
0.7222
31.99
10.99
Artistic
0.8776
0.9305
0.8753
0.8846
0.5370
30.75
12.42
Academic
0.8564
0.8518
0.8714
0.8481
0.2963
32.34
11.60
Average
0.8893
0.9105
0.8924
0.8876
0.5648
31.47
12.13
Table 4: Evaluation on the external MARIO-Eval benchmark. We report OCR-based text rendering metrics on a 500-sample external split. The main comparison is between each base model and its GRPO-aligned counterpart; Denotes the highest value in each column. model
NED
precision
recall
f1-score
accuracy
AnyText SD3.5 Qwen-Image FLUX FLUX(Our GRPO) Z-Image Z-Image(Our GRPO)
0.0896 0.3127 0.5861 0.2686 0.2974 0.5130 0.5189
0.0005 0.2621 0.5519 0.2253 0.2602 0.4849 0.4933
0.0008 0.3943 0.7236 0.3744 0.4185 0.6696 0.6852
0.0004 0.2816 0.5865 0.2437 0.2795 0.5199 0.5272
0.0000 0.1060 0.2860 0.1000 0.1240 0.2460 0.2600
signal around the hierarchical structure of rendering errors. By decomposing failures into global, word, and glyph levels, the VLM-based reward distinguishes missing text, word-level mismatches, and fine-grained character defects, and supports both GRPO and DPO. Experiments on FLUX.1-dev and Z-Image-Turbo show consistent gains in OCR-based text accuracy while largely preserving general image quality. Results on an external benchmark, category-wise analysis, robustness tests over text length and placement, qualitative comparisons, and human preference evaluation further indicate that careful reward design can improve reliable text rendering without model-specific architectural redesign.
9
Table 5: Ablation on the three reward levels. For each base model, we report the full TextAlign (all three levels) and three variants that remove the global, word, or glyph level respectively. model
NED
precision
recall
f1-score
accuracy
0.5758 0.5398 0.5417 0.5409 0.5301
0.5829 0.5387 0.5428 0.5356 0.5315
0.1926 0.1722 0.1648 0.1889 0.1667
0.8924 0.8826 0.8735 0.8762 0.8726
0.8876 0.8772 0.8740 0.8742 0.8657
0.5648 0.5371 0.5407 0.5167 0.5278
Base: FLUX FLUX(ours), full w/o global level w/o word level w/o glyph level FLUX (Baseline)
0.6075 0.5839 0.5877 0.5723 0.5768
0.6070 0.5799 0.5810 0.5786 0.5746 Base: Z-Image
Z-Image(ours), full w/o global level w/o word level w/o glyph level Z-Image (Baseline)
0.8893 0.8841 0.8998 0.8785 0.8739
0.9105 0.8984 0.8803 0.8961 0.8860
References [1] Black Forest Labs. FLUX.1. https://blackforestlabs.ai/, 2024. Text-to-image model suite and release documentation. [2] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. Textdiffuser: Diffusion models as text painters. In Advances in Neural Information Processing Systems, 2023. [3] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. In European Conference on Computer Vision, pages 386–402. Springer, 2024. [4] Q. Chen, Y. Ma, H. Wang, J. Yuan, W. Zhao, Q. Tian, H. Wang, S. Min, Q. Chen, and W. Liu. Infinitecanvas: Higher-resolution video outpainting with extensive content generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 2150–2158, 2025. [5] P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, K. Lacey, A. Goodwin, Y. Marek, and R. Rombach. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. [6] L. Gao, J.-Y. He, Y. Zeng, Y. Zhong, X. Sun, J. Hu, Z. Gao, and X. Wei. Vitype: High-fidelity visual text rendering via glyph-aware multimodal diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 4131–4139, 2026. [7] J. Hessel, A. Holtzman, M. Forbes, R. Le Bras, and Y. Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514–7528, 2021. [8] X. Hu, K. Xu, B. Liu, Q. Liu, and H. Fei. Amo sampler: Enhancing text rendering with overshooting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13157–13166, 2025. [9] F. Ji, J. Yang, Z. Song, L. Gao, J. Liang, Z. Chen, J. Zhang, and X. Chen. Servimage: An image generation and editing benchmark from real-world commercial imaging services, 2026. [10] F. Ji, J. Yang, Z. Song, Y. Wang, Z. Cui, Y. Li, Q. Jiang, and X. Chen. Finestate-bench: Benchmarking state-conditioned grounding for fine-grained gui state setting, 2026. [11] F. Ji, J. Yang, Z. Song, Y. Wang, Z. Cui, Y. Li, Q. Jiang, M. Fang, and X. Chen. Finestate-bench: A comprehensive benchmark for fine-grained state control in gui agents, 2025. [12] Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:36652–36663, 2023. [13] W. Kong, Q. Tian, Z. Zhang, R. Min, Z. Dai, J. Zhou, J. Xiong, X. Li, B. Wu, J. Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. [14] Z. Lai, Y. Zheng, Z. Cai, H. Lyu, J. Yang, H. Liang, Y. Hu, and B. Wang. Can multimodal llms see materials clearly? a multimodal benchmark on materials characterization. arXiv preprint arXiv:2509.09307, 2025.
10
[15] Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. In International Conference on Learning Representations, 2023. [16] R. Liu, D. Garrette, C. Saharia, W. Chan, A. Roberts, S. Narang, I. Blok, R. Mical, M. Norouzi, and N. Constant. Character-aware models improve visual text rendering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16270–16297, 2023. [17] X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In International Conference on Learning Representations, 2023. [18] Z. Liu, W. Liang, Z. Liang, C. Luo, J. Li, G. Huang, and Y. Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In European Conference on Computer Vision, pages 361–377. Springer, 2024. [19] J. Ma, M. Zhao, C. Chen, R. Wang, D. Niu, H. Lu, and X. Lin. Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation. arXiv preprint arXiv:2303.17870, 2023. [20] Y. Ma, H. Liu, H. Wang, H. Pan, Y. He, J. Yuan, A. Zeng, C. Cai, H.-Y. Shum, W. Liu, et al. Follow-youremoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. [21] Y. Ma, X. Wu, K. Chen, F. Zhu, R. Zhao, and H. Li. HPSv3: Towards wide-spectrum human preference score. arXiv preprint arXiv:2508.03789, 2025. [22] X. Meng, S. Huang, J. Yang, M. Ma, Z. Ma, L. Han, G. Yuan, H. Li, and L. Cheng. From reach to insert: Tactile-augmented precision assembly under sub-millimeter tolerances, 2026. [23] PaddlePaddle Team. PaddleOCR-VL: Boosting general document parsing via a 0.9B ultra-compact vision-language model. arXiv preprint arXiv:2510.14528, 2025. [24] W. Peebles and S. Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. [25] R. Rafailov, A. Sharma, E. Mitchell, S. Ermon, C. D. Manning, and C. Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, volume 36, 2023. [26] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. [27] C. Schuhmann. LAION-Aesthetics: A linear aesthetic quality predictor on top of CLIP embeddings. https://github.com/christophschuhmann/improved-aesthetic-predictor, 2022. [28] Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. K. Li, Y. Wu, and D. Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. [29] W. Shi, Y. Song, D. Zhang, J. Liu, and X. Zou. Fonts: Text rendering with typography and style controls. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18463–18474, 2025. [30] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. [31] Z. Song, J. Yang, Y. Huang, J. Tonglet, Z. Zhang, T. Cheng, M. Fang, I. Gurevych, and X. Chen. Geolocation with real human gameplay data: A large-scale dataset and human-like reasoning framework, 2026. [32] Ł. Staniszewski, B. Cywiński, F. Boenisch, K. Deja, and A. Dziedzic. Precise parameter localization for textual generation in diffusion models. In The Thirteenth International Conference on Learning Representations, 2025. [33] D. Tang, Q. Jiang, J. Yang, J. Zhao, X. Du, M. Fang, and X. Zhang. Sltp: A symbolic travel-planning agent framework with decoupled translation and heuristic tree search. Electronics, 15(2), 2026. [34] Y. Tuo, W. Xiang, J.-Y. He, Y. Geng, and X. Xie. Anytext: Multilingual visual text generation and editing. In International Conference on Learning Representations, 2024.
11
[35] B. Wallace, M. Dang, R. Rafailov, L. Zhou, A. Lou, S. Purushwalkam, S. Ermon, C. Xiong, S. Joty, and N. Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. [36] Y. Wang, C. Han, Y. Li, Z. Jin, X. Li, S. Du, W. Tao, S. Li, Y. Yang, C. Yuan, et al. Uniglyph: Unified segmentation-conditioned diffusion for precise visual text synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18335–18344, 2025. [37] Y. Wang, W. Zhang, H. Xu, and C. Jin. Dreamtext: High fidelity scene text synthesis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 28555–28563, 2025. [38] Z. Wang, J. Bao, S. Gu, D. Chen, W. Zhou, and H. Li. Designdiffusion: High-quality text-to-design image generation with diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20906–20915, 2025. [39] C. Wu, J. Li, J. Zhou, J. Lin, K. Gao, K. Yan, S.-m. Yin, S. Bai, X. Xu, Y. Chen, Y. Chen, Z. Tang, Z. Zhang, Z. Wang, A. Yang, B. Yu, C. Cheng, D. Liu, D. Li, H. Zhang, H. Meng, H. Wei, J. Ni, K. Chen, K. Cao, L. Peng, L. Qu, M. Wu, P. Wang, S. Yu, T. Wen, W. Feng, X. Xu, Y. Wang, Y. Zhang, Y. Zhu, Y. Wu, Y. Cai, and Z. Liu. Qwen-image technical report. arXiv preprint arXiv:2508.02324, 2025. [40] J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong. ImageReward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903–15935, 2023. [41] Z. Xue, J. Wu, Y. Gao, F. Kong, L. Zhu, M. Chen, Z. Liu, W. Liu, Q. Guo, W. Huang, and P. Luo. DanceGRPO: Unleashing GRPO on visual generation. arXiv preprint arXiv:2505.07818, 2025. [42] Z. Yan, J. Wang, A. Wang, Y. Li, W. Shang, and Z. Hangcheng. Textmaster: A unified framework for realistic text editing via glyph-style dual-control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16112–16121, 2025. [43] A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. [44] J. Yang, M. Cui, H. Zhang, F. Ji, Z. Lai, and Y. Wang. Agent-based anti-jamming techniques for uav communications in adversarial environments: A comprehensive survey, 2025. [45] J. Yang, Z. Han, M. Xiang, H. Wang, Y. Huang, and M. Fang. Asynchronous and segmented bidirectional encoding for NMT. CoRR, abs/2402.14849, 2024. [46] J. Yang, H. Wang, Q. Zhao, Z. Shi, Z. Song, and M. Fang. Efficient reinforcement learning via decoupling exploration and utilization. In International Conference on Intelligent Computing, pages 396–406. Springer, 2024. [47] J. Yang, H. Zhang, F. Ji, Y. Wang, M. Wang, Y. Luo, and W. Ding. Frequency point game environment for uavs via expert knowledge and large language model. Drones, 10(2), 2026. [48] Y. Yang, D. Gui, Y. Yuan, W. Liang, H. Ding, H. Hu, and K. Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36:44050–44066, 2023. [49] J. Yuan, X. Zhang, H. Zhou, J. Wang, Z. Qiu, Z. Shao, S. Zhang, S. Long, K. Kuang, K. Yao, et al. Hap: Structure-aware masked image modeling for human-centric perception. Advances in Neural Information Processing Systems, 36:50597–50616, 2023. [50] Z-Image Team, H. Cai, S. Cao, R. Du, P. Gao, S. Hoi, Z. Hou, S. Huang, D. Jiang, X. Jin, L. Li, Z. Li, Z.-Y. Li, D. Liu, D. Liu, J. Shi, Q. Wu, F. Yu, C. Zhang, S. Zhang, and S. Zhou. Z-Image: An efficient image generation foundation model with single-stream diffusion transformer. arXiv preprint arXiv:2511.22699, 2025. [51] B. Zhang, Z. Gao, Y. Qu, and H. Xie. How control information influences multilingual text image generation and editing? Advances in Neural Information Processing Systems, 37:6884–6904, 2024. [52] L. Zhang, X. Chen, Y. Wang, Y. Lu, and Y. Qiao. Brush your text: Synthesize any scene text on images via diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7215–7223, 2024. [53] Y. Zhao and Z. Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In European conference on computer vision, pages 217–233. Springer, 2024.
12
[54] Y. Zhu, J. Liu, F. Gao, W. Liu, X. Wang, P. Wang, F. Huang, C. Yao, and Z. Yang. Visual text generation in the wild. In European Conference on Computer Vision, pages 89–106. Springer, 2024.
A
Details of the Text Rendering Benchmark Construction
This section provides the full construction details summarized in Sec. 4.1. The pipeline is run independently per category and proceeds through five stages: target-text generation, CLIP-based text deduplication, prompt synthesis with explicit position control, CLIP-based prompt deduplication, and LLM-based content-safety filtering. We use Qwen3-8B [43] as the LLM backbone for all generation and classification steps, and CLIP ViT-B/32 as the text encoder for semantic deduplication. A.1
Category Taxonomy
We curate ten visual-text carriers that jointly cover the use cases observed in real deployment of textto-image models: Poster, Advertisement, Cover, Logo, Sticker, Handwriting, Scene, Basic, Artistic, and Academic. The categories are intentionally heterogeneous along four properties so that no single regime dominates the benchmark: Poster and Cover contain dense, multi-block compositions; Logo and Sticker use a single text block with strongly stylized typography; Handwriting and Artistic introduce the largest deviation from canonical glyph shapes; Scene and Academic present the most severe background interference at the smallest relative text scale; Advertisement and Basic cover commercial signage and informational layouts respectively. Each category is bound to a categoryspecific expert role and to a small pool of curated real-world reference samples, both of which act as stylistic anchors during text generation. A.2
Target-Text Generation (Length Axis)
For each category, we sample target texts by prompting Qwen3-8B with an expert-role system message tied to the category—e.g., a visual-communication and poster-design expert for Poster, a scientific-writing expert for Academic. The category-specific reference samples are inserted into the prompt purely for stylistic calibration; the model is explicitly instructed not to paraphrase them. To realize the text-length axis, generated texts are stratified into three buckets by word count—short (≤ 5 words), medium (6–15 words), and long (≥ 16 words)—and each bucket is sampled separately rather than relying on the LLM’s natural length prior, so that long-text samples are not crowded out by short ones. Sampling uses temperature 0.8 and top-p 0.9, with a batch size of 50 samples per LLM call. A.3
Prompt Synthesis (Complexity and Position Axes)
Given a target text y and its category, we issue a second LLM call to compose a single-paragraph image-generation instruction c that embeds y verbatim, enclosed in double quotation marks. The synthesis enforces two structural constraints, which jointly realize the remaining two axes of the benchmark. Position constraint (position axis). We control where the rendered text appears within the instruction by discretizing the relative character offset r = idx(y)/|c| of the quoted target into three positions: front (r < 1/3), middle (1/3 ≤ r < 2/3), and back (r ≥ 2/3). At generation time the LLM receives an explicit positional instruction. For the back bucket—where the LLM most frequently abandons the position constraint—we additionally apply a two-step assembly strategy: a scene description is first generated, and then concatenated with the quoted target via a connective phrase sampled from a curated pool (e.g., “displaying the text”, “with the inscription”, “inscribed with”). Each generated prompt is parsed back to compute its actual position, and prompts that violate the requested bucket are regenerated up to a fixed retry budget. Compositional content constraint (complexity axis). The prompt must specify, beyond the embedded text, three orthogonal visual aspects: scene content (subjects, objects, environment), spatial layout (placement and relative scale of the text block), and stylistic descriptors (lighting, palette, artistic style, material). To prevent collapse to short templated descriptions, prompts are themselves stratified by word count, excluding the rendered text, into short (≤ 15 words), medium (16–45 words), and long (≥ 46 words) buckets, each driven by a dedicated template that exposes a length-appropriate subset of the visual aspects above. Sampling uses temperature 0.85 and top-p 0.92. 13
A.4
CLIP-Based Semantic Deduplication
Even at high decoding temperature, LLM sampling produces near-duplicate outputs that bias the empirical distribution toward common templates. We therefore apply CLIP-based semantic deduplication twice: once after target-text generation, and once after prompt synthesis. Within each bucket, items are encoded into L2-normalized CLIP text embeddings, and we compute pairwise cosine similarity. For every item we record its maximum similarity to any other item in the bucket; we retain the bottom 80% in ascending order of this max-similarity score and discard the most redundant 20%. The same procedure is applied to prompts within each (category, prompt_length) sub-bucket. When the pipeline is run incrementally over multiple rounds, we additionally compute cross-similarity between new candidates and items already retained in prior rounds, and remove any new item whose maximum cross-similarity exceeds the retention threshold, so that the incremental rounds do not silently reintroduce duplicates. A.5
Content-Safety Filtering
Each surviving prompt is finally screened by Qwen3-8B operating as a content-safety classifier. The model evaluates the prompt against five explicit categories—sexual, violent, gore, hateful, and illegal content—and emits a structured JSON verdict of the form {is_nsfw, categories, reason}. Items flagged as policy-violating are removed. Low-entropy decoding (temperature 0.3, top-p 0.8) is used at this stage to stabilize the verdict. A.6
Dataset Schema and Splits
Each surviving sample is serialized as a JSONL record with fields index, text, prompt, class, text_length, prompt_length, and position, capturing the sample identifier, the target text, the full instruction, the category label, the text-length bucket, the prompt-length bucket, and the verified position of the target text within the prompt. These fields support both bucket-balanced sampling at training time and stratified evaluation along the category, length, and position axes, the latter being reported in Sec. 4.3 and Appendix B. Samples are partitioned into a training split and a held-out evaluation split for the experiments in Sec. 4.
B
Robustness to Text Length and Spatial Placement
To verify that TextAlign generalizes beyond the easy regime of short, centered text, we stratify the evaluation set along the annotated text_length and position fields and evaluate Z-Image-Turbo (Our GRPO) and FLUX (Our GRPO) independently on each subset. As shown in Fig. 5, the radar plot of Z-Image-Turbo forms a near-regular hexagon over the six length-position subsets, with NED stably falling in the range 0.87–0.92; Table 6 further shows that its NED on the long-text subset (0.8976) is even marginally higher than on the short-text subset (0.8650), and the three position buckets remain balanced (NED 0.87/0.92/0.87). FLUX, in contrast, exhibits a visibly irregular radar shape: NED collapses to 0.46–0.61 on the long-text subset, and the back-position exact accuracy drops to 0.1556, reflecting a systematic weakness of the base model on long and trailing-position text. These results indicate that TextAlign delivers consistent improvements across length-position combinations on the stronger backbone, rather than overfitting to a simplified subset of the benchmark.
C
Additional Qualitative Results across Visual Categories
Fig. 6 presents representative samples from Z-Image-Turbo (Our GRPO) across the ten visual-text categories, jointly covering the typical challenges of dense multi-block layouts (Poster, Cover), strongly stylized typography (Logo, Sticker), non-canonical glyph forms (Handwriting, Artistic), and small-scale text embedded in cluttered backgrounds (Scene, Academic). The aligned model produces legible, complete, and semantically correct text in every category while preserving category-specific visual idioms—the hierarchical layout of posters, the stroke continuity of handwriting, and the chalkboard-like materiality of academic scenes. This cross-category stability suggests that the global, word, and glyph levels of the hierarchical reward jointly cover the failure modes of distinct visual-text regimes, without requiring category-specific supervision signals.
14
Figure 5: Robustness to text length and spatial placement. Radar visualizations of FLUX (Our GRPO) and Z-Image-Turbo (Our GRPO) across text-length and position subsets.
Academic
Advertisement
Artistic
Basic
Cover
Handwriting
Logo
Poster
Scene
Sticker
Figure 6: Qualitative results across visual categories. Z-Image (Our GRPO) renders legible text across diverse visual text scenarios while preserving category-specific style and layout.
15
Table 6: OCR-based text rendering metrics grouped by target text length and requested spatial placement Model Category NED precision recall f1-score accuracy CLIPScore Aesthetics HPSv3 Short Middle Z-Image Long (Our GRPO) Front Middle Behind
0.8650 0.8917 0.8976 0.8694 0.9180 0.8669
0.8764 0.9379 0.9153 0.9206 0.9120 0.8972
0.8729 0.9038 0.8956 0.8774 0.9159 0.8789
0.8583 0.9036 0.9004 0.8784 0.9096 0.8742
0.7000 0.5889 0.3111 0.5444 0.5278 0.5278
30.83 32.00 31.58 33.51 30.64 30.26
4.966 4.998 4.938 4.886 4.992 5.025
11.96 12.33 12.10 11.56 12.41 12.41
Short Middle FLUX Long (Our GRPO) Front Middle Behind
0.7167 0.6053 0.4625 0.6085 0.6100 0.5663
0.6789 0.6111 0.4728 0.6429 0.5650 0.5549
0.6658 0.5657 0.4058 0.5782 0.5422 0.5168
0.6558 0.5650 0.4237 0.5869 0.5426 0.5151
0.4056 0.1278 0.0000 0.2000 0.1778 0.1556
30.05 30.08 28.82 30.73 29.05 29.16
5.278 5.286 5.256 5.202 5.308 5.310
12.31 11.98 11.35 11.62 12.12 11.90
16