Paper deep dive
GlyphBanana: Advancing Precise Text Rendering Through Agentic Workflows
Zexuan Yan, Jiarui Jin, Yue Ma, Shijian Wang, Jiahui Hu, Wenxiang Jiao, Yuan Lu, Linfeng Zhang
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/22/2026, 6:32:11 AM
Summary
GlyphBanana is a training-free agentic workflow designed to improve precise text and mathematical formula rendering in diffusion models. It utilizes a four-stage pipeline—extraction, draft preview, glyph injection (via frequency decomposition and attention re-weighting), and style refinement—to integrate system font rendering capabilities with generative flexibility. The paper also introduces GlyphBanana-Bench, a comprehensive benchmark for evaluating text rendering across diverse linguistic domains and complexity levels.
Entities (5)
Relation Signals (3)
GlyphBanana → introduces → GlyphBanana-Bench
confidence 100% · we introduce GlyphBanana, alongside a corresponding benchmark... GlyphBanana-Bench
GlyphBanana → uses → Frequency Decomposition
confidence 95% · GlyphBanana employs an agentic workflow that integrates auxiliary tools... Frequency Decomposition is used to strengthen the high frequency structure
GlyphBanana → appliesto → Diffusion Transformer
confidence 90% · Notably, our training-free approach can be seamlessly applied to various Text-to-Image (T2I) models
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Despite recent advances in generative models driving significant progress in text rendering, accurately generating complex text and mathematical formulas remains a formidable challenge. This difficulty primarily stems from the limited instruction-following capabilities of current models when encountering out-of-distribution prompts. To address this, we introduce GlyphBanana, alongside a corresponding benchmark specifically designed for rendering complex characters and formulas. GlyphBanana employs an agentic workflow that integrates auxiliary tools to inject glyph templates into both the latent space and attention maps, facilitating the iterative refinement of generated images. Notably, our training-free approach can be seamlessly applied to various Text-to-Image (T2I) models, achieving superior precision compared to existing baselines. Extensive experiments demonstrate the effectiveness of our proposed workflow. Associated code is publicly available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2603.12155v1
- Canonical: https://arxiv.org/abs/2603.12155v1
Trouble viewing inline? Open PDF directly →
Full Text
73,307 characters extracted from source content.
Expand or collapse full text
GlyphBanana: Advancing Precise Text Rendering Through Agentic Workflows Zexuan Yan 1,2∗ Jiarui Jin 2∗ Yue Ma 3 Shijian Wang 2,4 Jiahui Hu 5 Wenxiang Jiao 2 Yuan Lu 2† Linfeng Zhang 1† 1 Shanghai Jiao Tong University 2 Xiaohongshu Inc. 3 Hong Kong University of Science and Technology 4 Southeast University 5 South China University of Technology Figure 1. Gallery of various text rendering results sampled by GlyphBanana. Abstract Despite recent advances in generative models driving significant progress in text rendering, accurately gener- ating complex text and mathematical formulas remains a formidable challenge.This difficulty primarily stems from the limited instruction-following capabilities of cur- rent models when encountering out-of-distribution prompts. To address this, we introduce GlyphBanana, alongside a corresponding benchmark specifically designed for ren- dering complex characters and formulas. GlyphBanana ∗ Equal contribution. † Corresponding author. employs an agentic workflow that integrates auxiliary tools to inject glyph templates into both the latent space and attention maps, facilitating the iterative refinement of generated images.Notably, our training-free ap- proach can be seamlessly applied to various Text-to-Image (T2I) models, achieving superior precision compared to existing baselines.Extensive experiments demonstrate the effectiveness of our proposed workflow.Associated code is publicly available at https://github.com/ yuriYanZeXuan/GlyphBanana. 1 arXiv:2603.12155v1 [cs.CV] 12 Mar 2026 Out of Distribution (OOD), good in style but badin precision Highest precision, but with poor style Withindistribution, good in both precision and style Deterministic system font rendering. Rendering complex text with diffusion model when OOD. Rendering daily & common text with diffusion model. Bridge the Gap？ Figure 2. The illustration of motivation. We observe that while in-distribution cases show satisfying precision-style banlance, there exists huge gap between OOD cases and deterministic rendered texts. 1. Introduction Recent diffusion transformers [5, 29, 32, 33, 35, 37, 41] have demonstrated remarkable progress in image genera- tion, driving a wide range of applications such as com- mercial advertising, poster design, and scientific visualiza- tion. In these contexts, accurate text rendering plays a crit- ical role, imposing stringent demands on both the general- izability of diffusion models and their capacity for multi- lingual instruction following. Basic mainstreaming genera- tive models, such as Z-Image [49] and Qwen-Image [56], excel at rendering frequently encountered text, including short English phrases, common everyday Chinese expres- sions, and simple mathematical equations. But they perform poorly on rare English words, complex Chinese characters, and sophisticated scientific formulas (as exemplified in Fig- ure 2). To improve their precise text rendering performance, existing approaches can be broadly categorized into two paradigms, namely training-based and training-free meth- ods. Training-based approaches, such as GlyphByT5 [26] and FluxText [15], adopt strategies of either LoRA-based fine-tuning or fine-tuning on the text encoder. Despite their effectiveness in certain scenarios, these methods commonly suffer from limited generalization ability and a heavy re- liance on high-quality annotated datasets. Training-free methods, such as TextCrafter [48] and FreeText [66], typi- cally incorporate a glyph prior as a spatial layout constraint to regulate and guide text rendering. However, an overly strong glyph prior tends to disrupt the background and over- all visual style of the image, resulting in style inconsistency between the rendered text and its surrounding content. We also note that system font tools offer high-precision text ren- dering capabilities, yet lack flexibility, as they require hand- crafted designs to adapt to specific styles. In this paper, we propose a novel agentic workflow, termed GlyphBanana, which effectively integrates the pre- cise rendering capabilities of system font rendering tools with the generative flexibility of diffusion models, thereby enabling autonomous adaptation to arbitrary styles with- out requiring any manual design intervention. Specifically, GlyphBanana operates through the following four sequen- tial stages. In the extraction stage, GlyphBanana first em- ploys vision-language models to extract the target text con- tent and the desired rendering style from the input prompt. Subsequently, in the draft preview stage, text-to-image models are applied to generate a preliminary image in the desired style as a reference preview, which is followed by a layout planner equipped with text grounding tools to produce a glyph template that encapsulates detailed at- tributes, including font type, color, bounding box coordi- nates, and rotation parameters. The glyph injection stage constitutes the core component of GlyphBanana, wherein the produced glyph template is integrated into the genera- tive model through both latent space and attention modules. Specifically, for the latent space, frequency decomposition is employed to disentangle the denoising representations of the glyph template into low- and high-frequency com- 2 A text book page displays“PV=nRT”in elegant font. Image Tokens Text Tokens Frequency Decomposition Txt2Img Extraction Input Text Extract style and text to refine. Tex t: <PV=nRT> Style: elegant font, researchpaper style Draft Preview Draft Image Analyze Font, Color, Grounding text and rotation parameters. Txt2Img Font: x, Glyph: Color: #hex BBox:[[x1,y1,x2,y2],] Rotation:20° Glyph Injection LatentSpace Img2Img Style Refinement Style Target Bold SizeItalic "text style": ”text printed in book page... "colors”: #hex Keep background unedited and make foreground text harmonize... Image Score Judge Tool Bank -Glyph Template-Diffusion Model DiT Block N ... DiT Block 1 DiT Block N 푉 ... 푷 풔풓풄 푍 ! “퐴n 푒푚푝푡푦 푡푒푥푡 푏표푘 푝푎푔푒.” VAE Encoder 푋~풩(휇,휎 ! ) concat 푍 !"# DiT Block 1 퐾 푄 img2img img2txt txt2img txt2txt Re-weight ... 푄 * Re-weight 푉 퐾 - 푍 !"$ 푍 !"$"# ... VAE 푍 % Font Controller Layout Planner Condition Extractor Style Refiner Denoising with Attention Re-weighting Injection -Attention Controller-Text Grounding Formula -Font Style-Formula Renderer Iterative Refine · Softmax( ) Attention Module DiT Block N ... DiT Block 1 Injection inTransitional Phase Figure 3. Overview of the GlyphBanana agentic pipeline. The workflow comprises four stages: (1) Extraction Stage parses the input into text content and style attributes; (2) Draft Preview Stage generates an initial image via a Layout Planner; (3) Glyph Injection Stage applies Frequency Decomposition in latent space and Attention Re-weighting inside each DiT block; (4) Style Refinement Stage employs iterative refinement with a Style Refiner and Score Judger. The bottom panel details the denoising process with the Attention Re-weighting. ponents, after which the information-dense high-frequency components are injected into the latent space. For the atten- tion modules, an attention re-weighting mechanism is in- troduced to incorporate the glyph template as a bias term into the attention maps within each DiT block. Finally, in the style refinement stage, the intermediate image gener- ated from the glyph injection stage are iteratively refined by jointly optimizing the refinement prompts and the generated images to further enhance overall image quality. It is worth noting that GlyphBanana is a training-free framework or- chestrated by a collection of plug-and-play tools, enabling seamless integration with arbitrary generative models. Existing text-rendering benchmarks [7, 24, 28, 48, 51], are narrowly focused on common English words or Chi- nese characters, systematically neglecting rare characters and complex scientific formulas. To address this limita- tion, we introduce GlyphBanana-Bench, a comprehensive text-rendering benchmark that, to the best of our knowl- edge, is the first to systematically evaluate text rendering across a diverse spectrum of difficulty levels and linguis- tic domains, ranging from simple common words and rare Chinese characters to complex multiline scientific formu- las. GlyphBanana-Bench is constructed through a com- bination of community-forum crawling and synthesis via Kimi-K2.5 [50], ensuring both diversity and scalability of the benchmark data. Extensive experiments demonstrate that our GlyphBanana achieves substantial improvements in OCR accuracy, attaining scores of 85.9 (+19.6%) on Z- Image and 75.8 (+6.91%) on Qwen-Image, while simulta- neously enhancing precision and style. 2. Related work DiT for Image Generation& Editing. Diffusion Trans- former (DiT) [41] has emerged as an alternative to U- Net [43] for image generation and editing. Building upon this, recent works [2, 4–7, 14, 25, 27, 30, 34, 36, 45– 47, 49, 53–56, 58, 62–64, 69] integrate Flow Match- ing [22] to improve training stability and inference ef- 3 ficiency.Benefiting from its unified attention architec- ture, DiT also demonstrates strong capabilities in image editing. Existing approaches can be broadly categorized into single-turn and multi-turn paradigms.Single-turn methods, such as GLIDE[38], MagicBrush[65], Prompt- to-Prompt[10], UltraEdit[9], and FireEdit[68], perform instruction-guided edits in a one-shot manner. In contrast, multi-turn systems[8, 39] enable iterative, context-aware editing through interactive feedback. Visual Text Rendering. Although diffusion-based meth- ods can generate high-quality images, rendering text in images remains a challenging problem due to the need for accurate spelling, layout coherence, and style consis- tency. One line of work[1, 24, 44] leverages large language models[42, 57] to improve spelling accuracy in generative models. Another line[3, 28, 60] focuses on explicitly con- trolling text layout and content during generation. Recent works further improve rendering quality from multiple per- spectives. TextCenGen[17] and TextCrafter[48] enhance layout and attribute consistency, while Calligrapher[31] and TextMaster[59] focus on style control via glyph- and feature-level guidance.SceneVTG[71] adopts a plan- ning–rendering pipeline with Vision Language Models to ensure semantically coherent text. Image Rendering with Agentic Workflow.Beyond single-step generation, real-world design tasks [12] of- ten require multi-step reasoning, iterative refinement, and human-like decision making. PosterGen[67] simulates a design team with specialized agents for layout and styling, Agent Banana[61] proposes a hierarchical planner-executor framework with long-horizon memory and layer-wise ma- nipulation.For image and video restoration, MoA-VR and AgenticIR[23, 70] extend agentic workflows to VLM- integrated multi-agent repair frameworks. In more complex settings such as creative photo retouching and task-oriented restoration, systems like JarvisIR, JarvisArt, 4KAgent, and JarvisEvo[18–20, 72] further demonstrate the effectiveness of agentic pipelines. Complementary to these system-level designs, EditThinker [16] focuses on enhancing intra-agent capability by formulating image editing as an explicit itera- tive reasoning process. 3. Preliminaries 3.1. Multimodal Diffusion Transformer The Multimodal Diffusion Transformer (M-DiT) mod- els the generation of an image I conditioned on a text prompt P within a latent space. First, a pre-trained Varia- tional Autoencoder (VAE) compresses the image into a low- dimensional latent representation z 0 = VAE enc (I). Following standard diffusion models, a forward process gradually corrupts the data z 0 into Gaussian noise by adding noise ε ∼ N(0, I). The noisy latent z t at timestep t is de- fined as: z t = α t z 0 + σ t ε,(1) where α t and σ t are the noise schedule parameters. The training objective J is to learn a neural network, parame- terized by θ, to reverse this process by predicting the added noise: minJ θ = minE t,z 0 ,ε h ∥ε θ (z t ,t,P)− ε∥ 2 2 i .(2) To parameterize the denoiser ε θ (z t ,t,P) using the M- DiT architecture, the continuous latent states and discrete text condition must be transformed into sequence rep- resentations.The noisy latent z t is spatially patchified and linearly projected to form the image tokens X img = Patchify(z t ). Simultaneously, the text prompt P is mapped by a pre-trained text encoder into a sequence of text tokens X txt = TextEnc(P). The two modality-specific token se- quences are then concatenated along the sequence dimen- sion to construct the joint hidden state for the Transformer network: H = [ X img ∥ X txt ].(3) After passing through the stacked M-DiT blocks, the updated visual components of H are separated and un- patchified back to the original spatial shape to yield the final noise prediction ε θ . 4. Methods As illustrated in Fig. 3, our agentic workflow comprises four tightly coordinated stages: (1) Extraction, which parses the user input into text content and style attributes; (2) Draft Preview, which generates a preliminary image and derives a typography plan; (3) Glyph Injection, which integrates pre- cise glyph information via Frequency Decomposition and Attention Re-weighting; and (4) Style Refinement, which iteratively improves visual harmony. The injection proce- dure is formalised in Algorithm 1. We detail each stage below. 4.1. Extraction Stage Given the user prompt P user , an extractor decomposes it into two components: the target text content T to be rendered, and a style description S that characterises the desired visual appearance.This stage output reference ground-truth for identifying text to be rendered in subse- quent stages. 4.2. Draft Preview Stage The draft preview stage produces a preliminary image and a detailed typography plan that guides glyph injection. In this stage, a draft image I draft is generated according to the original prompt P user , which is analysed by a Layout 4 Category 1: General Text for Rendering “category”: “F-Complex-L1”, “length”: “medium”, “prompt”: ‘A study guide with “∮∂V F·dS=∫V ∇·FdV”, “∮∂S F·dl=∫S (∇×F)·dS” wr itten aca dem ically.’, “text”: [" ∮ ∂V F·dS=∫V ∇ ·FdV", " ∮ ∂S F·dl=∫S ( ∇ ×F)·dS"], “text_length”: 37, “data_id”: 37, “lines”: 2 Category 2: Formulas from Easy to Complex RareEn. Easy En. Example: F-Medium F-Complex-L1 F-Complex-L2 F-Easy Auxiliary Tools: 1.ℳ ! :풯 "#" →풯 퓉ℯ퓍 : Convert text to formula which can be rendered by MathJaxengine, output a re-written prompt. 2.퐹 '(")*(# (⋅) : Convert Tex code to formula, output SVG. 3.Print (⋅) : Render SVG on empty canvas with given color, size and location, output a standard image. 4.Segment(⋅) : Apply Otsu to extract formula’s glyph mask from printed image, output a mask image. 5.풢(⋅): Grounding text or formula in image, analyze layout, font, color and glyph size, output a layout config in jsonl. RareZh. Easy Zh. Artistic letter Example: “category”: “Easy En.”, “length”: “short”, “prompt”: ‘A color chart displays “Orange” in bright citrus.’, “text”: ["Orange"], “text_length”: 6, “data_id”: 49, Optional: [“ref_image”:”path_to_ref_img”, “mask_image”: “path_to_mask”] Auxiliary Tools: 1.풜: Loading existing font library, output font collection. 2.Segment(⋅) : Apply Otsu to extract text mask from rendered glyph template, output a mask image. 3.풢(⋅): Grounding text or formula in image, analyze layout, font, color and glyph size, output a layout config in jsonl. Benchmark Figure 4. Illustration of the GlyphBanana-Benchmark with auxiliary tools. The proposed benchmark consists of two categories. General Text for Rendering assesses standard and stylized text rendering. Formulas from Easy to Complex evaluates formula rendering across varying complexities Table 1. Comparison of different text-rendering datasets. Num. refers to the number of samples in the dataset, and Avg.L refers to the average length of the char to be rendered in the dataset. refers to FLUX.2-klein-9B andrefers to Qwen-Image-2512. Datasets Text TypeConditionStatisticsOCR ScoreStyle Score En.Zh.FormulasImageMaskNum.Avg.L DrawTextExt [24]✓✗22017.00.760.810.830.80 AnyText [51]✓✗1000 21.80.330.440.670.69 CVTG-2K [48]✓✗2000 39.50.490.510.750.67 LongText-Bench [7]✓✗320 116.70.380.720.690.74 Ours✓29032.7 0.37 0.71 0.68 0.71 Planner, powered by the VLM equipped with text ground- ing tools. The planner creates a typography plan detailing the font, color, bounding boxes, and rotation angles for the generated text. This information is forwarded to the next stage to construct the injection template. 4.3. Glyph Injection Stage Formula Renderer produces pixel-accurate glyph images via L A T E X compilation for formulas, while for regular text, a Font Controller selects the appropriate font family, weight, and size according to typography plan P from last stage, outputing accurate glyph template image I . Along with rendered formulas, they are encoded by VAE into a strong glyph prior z tpl as a template. 4.3.1. Frequency Decomposition. Frequency Decomposition is used to strengthen the high frequency structure of the glyph template in denoising la- tent by precisely injecting high frequency glyph details. We define the frequency-decomposed blending function as fol- lows: F F.D. (z,z tpl ,M) = LF(z)+HF(z)⊙(1−M)+HF(z tpl )⊙M, (4) where LF(z)= GaussianBlur(z) extracts the low frequency component, HF(z)=z− LF(z) extracts the high frequency residual, and M is the mask that specifies the glyph-covered tokens. Specifically, GaussianBlur is implemented by a Gaussian blur kernel to do average pooling in latent space on image which is rendered using system font by font con- troller agent according to the typography plan. For the mask M , we use Otsu’s [40] method to segment the im- age into foreground and background. Since directly inject- ing the glyph latent into the denoising latent may lead to artifacts, we set a injection window [τ start ,τ end ) to con- trol the injection timing, leaving space for adjusting edge smoothness and style consistency with the background for diffusion model. This part is applied at each denoising step t∈ [τ start ,τ end ) in the Algorithm 1, stage 3. 5 Algorithm 1 Injection with Attention Enhancement Require: Typography planP , Prompt T , Total steps N , Injection window [τ start ,τ end ), Bias scales (0 < s − < 1 < s + ). Ensure: Glyph-injected latent z 0 . 1: I ← FontRender(P); M ← Otsu(I) ▷ Stage 1: Preprocessing 2: I txt ← FindTokenIndices(T, quoted) 3: I img ←i| M[i] > 0; ̃ I img ←i| M[i] = 0 ▷ glyph / non-glyph indices 4: ̃z list : ̃z 0 ,..., ̃z N ← Inversion(VAE(I)) ▷ fusion glyph template list 5: for denoising step t = N,..., 1 do 6:for AttnProcessor i ∈ DiT Block do ▷ Stage 2: Attn. re-weighting 7:B ←0; α + ← log(s + ); α − ← log(s − ) 8:B[I img ,I txt ] += α + ; B[I txt ,I img ] += α + ▷ enhance 9:B[ ̃ I img ,I txt ] += α − ; B[I txt , ̃ I img ] += α − ▷ suppress 10:Q,K,V ← Linear(h i,t ); ˆ Q, ˆ K ← RoPE(Q,K) 11:h i,t−1 ← SDPAttention( ˆ Q, ˆ K,V, bias=B) ▷ SDPAttention definition Eq. (5) 12:end for 13:if t/N ∈ [τ start ,τ end ) then ▷ Stage 3: Latent Injection 14: ̃z tpl ← ̃z list [t] 15:z t ←F F.D. (z t+1 , ̃z tpl , M) ▷ Frequency Decomposition Eq. (4) 16:end if 17: z t−1 ← Scheduler.step(z t ,t) 18: end for 19: return z 0 4.3.2. Injection with Attention Enhancement. We introduce a technique called Glyph Injection to in- ject the glyph latent into the denoising latent as shown in the Algorithm 1, stage 2. As inspired by P2P [10], TextCrafter [48], manipulating the attention value in the attention processors of DiT blocks could effectively af- fect response of output response pattern to prompt tokens. Specifically, we inject the Self-Attention module within the DiT block. Following the standard attention formulation in Transformers [52], the computation incorporating a bias matrix B is expressed as follows: SDPAttention( ˆ Q, ˆ K,V,B) = softmax ˆ Q ˆ K ⊤ √ d + B ! V, (5) where ˆ Q, ˆ K, and V denote the projected query, key, and value matrices respectively. d is the scaling dimension, and B is the attention bias matrix designed for explicit re- weighting. The matrix B is initialized to zero, and its non- zero elements B i,j are assigned by glyph-template in latent space, illustrated in Fig. 3, stage 3. We use the glyph latent template ̃z tpl to specify image tokens that are likely to be af- fected by the glyph injection, which can be divided into two parts: glyph-covered tokens, defined by Indices I img and non-glyph-covered tokens, defined by Indices ̃ I img . Simi- larly, text tokens are extracted and its corresponding indices are defined asI txt . For precisely control the attention com- puting process, we enhance the attention map value from I img to I txt and suppress value from I txt to ̃ I img in the attention processors of DiT blocks. 4.4. Style Refinement Stage 4.4.1. Iterative Refinement. To improve text rendering quality and ensure stylistic har- mony with the background, we introduce an Iterative Re- fine module. As illustrated in Fig. 3, this module utilizes a pretrained image-to-image diffusion model F DM to re- fine the output of the Glyph Inject stage. The refinement is driven by a VLM that serves dual functions: a Style Re- finer that identifies and corrects discordant visual attributes (e.g., color, texture, shadow) based on intermediate outputs and produces an amended prompt P ′ , and a Score Judger that evaluates each candidate and selects the optimal result. Formally, given the injected image I origin , its typogra- phy plan prompt P , and the glyph mask M , we construct a diverse candidate pool from three refinement strategies: I mask = M ⊙ I origin + (1−M)⊙F DM (I origin | P) which restricts regeneration to the non-glyph region to preserve text contours, I ref = F DM (I origin | P, M) which condi- tions on M as a reference to guide generation while allow- ing broader stylistic adjustment, and I sty = F DM (I origin | P ′ , M) where P ′ is the amended prompt produced by the Style Refiner. The Score Judger then selects the best output from the candidate pool: I ∗ = argmax I∈I origin ,I mask ,I ref ,I sty S VLM (I,P),(6) where S VLM denotes the VLM-based quality assessment. The system operates in a closed loop: the Style Refiner an- alyzes I ∗ , updates P ′ , regenerates the candidate pool, and the Score Judger re-evaluates, iterating until convergence or a maximum number of rounds is reached. 5. Benchmark and Evaluation Protocals 5.1. Benchmark Current evaluation frameworks for text-rendering diffusion models inadequately assess out-of-vocabulary (OOV) to- kens, complex notation, and the hierarchical multiline lay- outs typical of scientific equations. To bridge this evaluation gap, we present the GlyphBanana-Benchmark as illustrated in Fig. 4. We delicately collect and construct a wide range of text and formulas to be rendered along with supplemen- tary tools for agentic workflow. For category of general text, 6 A traditional shop sign shows "Halcyon" in golden letters. 用温馨棕色字体写着 "床铺" A research paper displays "PV=nRT" in elegant LaTeX font. A research document with "w+x+y+z=10", "w-x+y-z=2", "2w+x- y+z=7", "w+2x+3y- z=8" in formal notation. Ours Flux-Klein GLM-image Flux-dev Te x t C ra f t e rAnyText2 FailFail 一块⻛化的招牌上用 传统书法写着"遮莫 （不管，任凭）" FailFail A beach towel shows "Sand" in warm tan letters. Aphysics lecture board shows "C ₆ H ₁₂ O ₆ +6O ₂ →6CO ₂ +6H ₂ O" written in white chalk. Figure 5. Qualitative comparisons with other baselines. Fail denotes the FLUX.1-dev based models unable to follow instructions to render chinese text due to its limited text-encoder. Besides, we color the quoted text in red, referring to the target text to be rendered, and color the style text related to the glyph in blue. we provide auxiliary tools including font library, segmenta- tion and text-grounding tools for specifying text font and layout. To the best of our knowledge, it is the first bench- mark to systematically evaluate text rendering capabilities across a comprehensive difficulty spectrum ranging from simple words to complex, multiline mathematical formu- las, while supporting multimodal inputs and auxiliary ren- dering tools. The dataset is meticulously constructed: the rare Chinese word subset is curated by crawling commu- nity forums [13], whereas the English and complex formula subsets are entirely synthesized using Kimi-K2.5 [50]. Furthermore, we conduct quantitative evaluations with other similar benchmark on text type, input conditions, statistics of benchmark size, and score related to preci- sion and style metrics. Specifically, we employ two pop- ular open-source diffusion models to assess existing base- line metrics, which are FLUX.2-klein-9B andQwen- Image-2512 in Table. 1.Results reveal that accurately rendering rare Chinese characters and complex formulas remains a challenge for current diffusion-based methods. More qualitative results refer to supplementary materials. 5.2. Evaluation Protocols We adopt a multi-dimensional evaluation protocol cover- ing Optical Character Recognition(OCR) Score, Vision- Language Model(VLM) Score, Image-Text Matching(ITM) Score, and User Study. OCR Score represents the preci- sion of the rendered text, here we use OCR Accuracy(OCR- 7 Table 2. Quantitative comparison results for text-rendering metrics. Method OCR Score↑VLM Score↑ITM Score↑User Study↓ Acc.Ned.StyleFaith.VQACLIP Aesthetic Faith. Ours +Z−Image w/o re-weight72.376.70.7450.7040.8080.7092.732.70 w/o refine84.086.70.7250.7450.7980.7103.473.40 w/o F.D.84.587.30.7550.7640.8030.7082.732.83 full85.988.10.7650.7640.8140.7201.071.07 Ours +QwenImage w/o re-weight70.774.80.6760.6890.8140.6802.432.23 w/o refine75.779.60.6870.7770.8200.6973.603.43 w/o F.D.74.578.70.7180.8120.8190.6962.933.00 full75.879.90.7290.8300.8390.6941.031.33 Table 3. Quantitative comparison results for text-rendering metrics.represents for FLUX.2-klein-9B andrepresents for Qwen- Image-2512. Method OCR Score↑VLM Score↑ITM Score↑User Study↓ Acc.Ned.StyleFaith.VQACLIPAesthetic Faith. AnyText233.840.50.6610.4380.6410.6377.807.62 TextCrafter34.039.60.6720.3710.8040.6806.756.47 Flux.1 .dev27.934.30.6910.2800.7710.6396.757.03 FluxText25.028.40.6000.3510.7180.6566.986.83 Flux.2 .klein 36.742.10.6760.5210.8210.6866.386.08 GLM-Image62.170.80.7280.7000.8070.6815.505.77 Zimage71.876.30.7500.7030.8130.7235.075.12 Qwen-Image70.974.70.7050.7670.8400.6994.634.98 Ours +zimage 85.9↑14.188.1↑11.80.765↑0.015 0.764↑0.061 0.814↑0.0010.720↓0.0032.272.58 Ours +QwenImage 75.8↑4.979.9↑5.20.729↑0.0240.830↑0.0630.839↓0.001 0.694↓0.0052.872.52 Acc) and OCR Normalized Edit Distance(OCR-NED) to evaluate the precision of the rendered text. Define d(g,p) as the Levenshtein distance between the ground-truth text g and the predicted text p. OCR-Acc = 1 − d(g,p)/|g|, is a recall-oriented score that quantifies how much of the ground-truth text g is correctly rendered in the prediction p. OCR-NED = 1− d(g,p)/ max(|g|,|p|), is a symmet- ric similarity that additionally penalizes hallucinated text. Following recent practice, we additionally query a VLM to obtain VLM Score, including VLM-Style related to clarity, coherence, and aesthetics and VLM-Faithfulness related to scene, object, style, and text placement adherence to the prompt. Image-Text Matching measures the alignment between the rendered text and the reference image, using CLIP Score [11] and VQA Score [21]. User Study is eval- uated by performing a human preference sheet on aesthetic and faithfulness preferences to evaluate the text rendering quality, ordering images from best to worst. 6. Experiments 6.1. Implementation Details In our experiments, For Text-to-Image Generation, we adopt two open-source diffusion backbones: QwenImage- 2512 with 50 denoising steps and ZImage-turbo with 20 de- noising steps. For Style-Refiner, FLUX.2-klein-9B model is adopted for Image-to-Image Generation. To support the agent’s core planning and evaluation, we employ the pow- erful open-source Qwen3-VL-235B-A22B-Instruct model, which serves as the Layout Planner during the Draft Pre- view stage, Style Refiner and Score Judger during the Style Refine stage, OCR excecutor and VLM Score eval- uator for evaluation. All experiments are conducted on NVIDIA H800 GPUs. During Glyph Injection, attention 8 一张神秘的海报上用发光字 体显示"玓瓑(珠光闪耀)" A lecture board displaying "Av=λv" A problem set showing "(iγᵘ∂ᵤ-m)ψ=0", "γᵘ,γᵛ=2gᵘᵛ", "u,v=0,1,2,3" with equations. w/o Injectionw/o Refine Iterative Refine Figure 6. Qualitative comparison results. re-weighting is applied over denoising timesteps in the bias of scaled-dot-production-attention in the range (0.2, 0.8), with 2.0 enhancement scale and 0.1 for attention suppres- sion. More details are provided in the supplementary mate- rials. 6.2. Comparison with baselines Qualitative Comparison. We conducted extensive experi- ments with our proposed Agentic method (Z-Image by de- fault) on GlyphBanana-Benchmark, comparing it with other approaches. Specifically, we comprehensively compare the text rendering capabilities of the models on simple to rare English, Chinese, and from simple to complex multiline for- mulas. As shown in Figure 5, our method achieves supe- rior performance in text rendering on simple to rare En- glish, Chinese, and from simple to complex multiline for- mulas compared to existing methods. For FLUX.1-dev and TextCrafter, Chinese rendering is not supported due to its limited text-encoder. For multiline formulas, other methods show lower precision or duplicate rendering. Our method supports multi-language, and shows highest precision for rendering formulas. Quantitative Comparison. We performed comprehensive quantitative experiments on GlyphBanana-Benchmark. Our methods significantly improve the metrics related to ren- dering precision and quality score, including User Study, achieving the highest text accuracy among all other methods including text-rendering specific approaches such as Any- Text2 and TextCrafter. Compared to the original text ren- dering baseline, the T2I matching metrics are nearly identi- cal, but the style and faithfulness scores are higher including 7.62 7. 8 8 8.25 9.12 4.88 5.77 8.15 8.19 85.9 75.8 +Z-Image+Qwen-Image w/o refinew/o refine Iter. 1 Iter. 1 Iter. 2Iter. 2Iter. 3 Iter. 3 Tex t A c. Visual Quality Figure 7. Metric comparisons for multi-turn refinement. w/o F.D.w/ F.D.w/o F.D.w/ F.D. Figure 8. Qualitative comparisons for illustrating methods of Fre- quency Decomposition. User Study, getting the best overall performance. 6.3. Ablation Study Extensive qualitative and quantitative experiments are con- ducted using text accuracy, image quality, and user study metrics to validate the effectiveness of three key operations in our agentic workflow: Frequency Decomposition (F.D. for short), Attention Enhancement(re-weight for short), and Iterative Refine(refine for short). Ablation study of F.D. in latent space. Fig. 8 shows the impact of F.D. for improving the text rendering quality. 9 During rendering, the unwanted dark edges persist along- side the text strokes, as can be seen in the contours of ’Royal’ and ’Magi’ without F.D.. With F.D., the text is ren- dered more harmoniously with the background than with- out F.D. It illustrates that F.D. preserves space for style and color while maintaining the text structure. In addition, Ta- ble 2 demonstrates metrics across precision faces compre- hensive decline without F.D.. Ablation study of Injection. It can be observed from Fig. 6 that the glyph injection significantly improves the text pre- cision, which can be verified by OCR scores shown in Ta- ble 2. It demonstrates that leveraging the glyph information to re-weight attention value, and injecting the glyph latent into the latent space gain significant improvement on text rendering precision. Ablation study of Iterative Refine. The iterative refine process is shown in the right side of Fig. 6, which signif- icantly improves the text rendering quality, and trends can be visual by the Fig. 7. This process contributes to improv- ing style score without harming the rendering accuracy. It indicates that iterative refinement steadily enhances the Vi- sual Quality of the rendered text while largely preserving Text Accuracy, demonstrating the effectiveness of our Style Refinement. 7. Conclusion We present GlyphBanana, a training-free agentic frame- work that bridges font-level precision and diffusion-model flexibility via frequency-decomposed latent injection, at- tention re-weighting, and VLM-driven iterative refinement. Without any fine-tuning, it generalises across DiT back- bones and surpasses all baselines in both rendering accuracy and visual quality. We further contribute GlyphBanana- Bench, the first benchmark covering common words, rare characters, and complex scientific formulas. References [1] Y. Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, M. Aittala, Timo Aila, S. Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensem- ble of expert denoisers. ArXiv, abs/2211.01324, 2022. 4 [2] Siyu Cao, Hangting Chen, Peng Chen, Yiji Cheng, Yutao Cui, Xinchi Deng, Ying Dong, Kipper Gong, Tianpeng Gu, Xiusen Gu, et al. Hunyuanimage 3.0 technical report. arXiv preprint arXiv:2509.23951, 2025. 3 [3] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. ArXiv, abs/2305.10855, 2023. 4 [4] Yiyang Chen, Xuanhua He, Xiujun Ma, and Yue Ma. Con- textflow: Training-free video object editing via adaptive con- text enrichment. arXiv preprint arXiv:2509.17818, 2025. 3 [5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ̈ uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yan- nik Marek, and Robin Rombach. Scaling rectified flow trans- formers for high-resolution image synthesis, 2024. 2 [6] Kunyu Feng, Yue Ma, Bingyuan Wang, Chenyang Qi, Haozhe Chen, Qifeng Chen, and Zeyu Wang. Dit4edit: Dif- fusion transformer for image editing.In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2969– 2977, 2025. [7] Zigang Geng, Yibing Wang, Yeyao Ma, Chen Li, Yongming Rao, Shuyang Gu, Zhao Zhong, Qinglin Lu, Han Hu, Xi- aosong Zhang, Linus, Di Wang, and Jie Jiang. X-omni: Reinforcement learning makes discrete autoregressive image generative models great again, 2025. 3, 5 [8] Google DeepMind.Gemini 2.5 flash image (nano ba- nana). https://aistudio.google.com/models/ gemini-2-5-flash-image, 2025. Accessed: 2025- 10-29. 4 [9] Xiaojie Gu, Ziying Huang, Jia-Chen Gu, and Kai Zhang. Ul- traedit: Training-, subject-, and memory-free lifelong editing in language models, 2025. 4 [10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. 2022. 4, 6 [11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met- ric for image captioning, 2022. 8 [12] Yuxin Jiang, Yuchao Gu, Yiren Song, Ivor Tsang, and Mike Zheng Shou. Personalized vision via visual in-context learning. arXiv preprint arXiv:2509.25172, 2025. 4 [13] Jingluohaidijiwanli. How many rare characters are there in chinese?, 2020. Zhihu Answer. 7 [14] Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, Sumith Ku- lal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas M ̈ uller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space, 2025. 3 [15] Rui Lan, Yancheng Bai, Xu Duan, Mingxing Li, Dongyang Jin, Ryan Xu, Dong Nie, Lei Sun, and Xiangxiang Chu. Flux-text: A simple and advanced diffusion trans- former baseline for scene text editing.arXiv preprint arXiv:2505.03329, 2025. 2 [16] Hongyu Li, Manyuan Zhang, Dian Zheng, Ziyu Guo, Yi- meng Jia, Kaituo Feng, Hao Yu, Yexin Liu, Yan Feng, Peng Pei, Xunliang Cai, Linjiang Huang, Hongsheng Li, and Si Liu. Editthinker: Unlocking iterative reasoning for any im- age editor, 2025. 4 [17] Tianyi Liang, Jiangqi Liu, Yifei Huang, Shiqi Jiang, Jian- shen Shi, Changbo Wang, and Chenhui Li.Textcen- gen: Attention-guided text-centric background adaptation for text-to-image generation, 2025. 4 [18] Yunlong Lin, Zixu Lin, Haoyu Chen, Panwang Pan, Chenxin Li, Sixiang Chen, Wen Kairun, Yeying Jin, Wenbo Li, and Xinghao Ding. Jarvisir: Elevating autonomous driving per- ception with intelligent image restoration. In Proceedings of 10 the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 4 [19] Yunlong Lin, Zixu Lin, Kunjie Lin, Jinbin Bai, Panwang Pan, Chenxin Li, Haoyu Chen, Zhongdao Wang, Xinghao Ding, Wenbo Li, and Shuicheng Yan. Jarvisart: Liberating human artistic creativity via an intelligent photo retouching agent. arXiv preprint arXiv:2506.17612, 2025. [20] Yunlong Lin, Linqing Wang, Kunjie Lin, Zixu Lin, Kaix- iong Gong, Wenbo Li, Bin Lin, Zhenxi Li, Shiyi Zhang, Yuyang Peng, Wenxun Dai, Xinghao Ding, Chunyu Wang, and Qinglin Lu. Jarvisevo: Towards a self-evolving photo editing agent with synergistic editor-evaluator optimization, 2025. 4 [21] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text gen- eration. arXiv preprint arXiv:2404.01291, 2024. 8 [22] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxi- milian Nickel, and Matt Le. Flow matching for generative modeling, 2023. 3 [23] Lu Liu, Chunlei Cai, Shaocheng Shen, Jianfeng Liang, Weimin Ouyang, Tianxiao Ye, Jian Mao, Huiyu Duan, Jiangchao Yao, Xiaoyun Zhang, Qiang Hu, and Guangtao Zhai. Moa-vr: A mixture-of-agents system towards all-in- one video restoration, 2025. 4 [24] Rosanne Liu, Dan Garrette, Chitwan Saharia, William Chan, Adam Roberts, Sharan Narang, Irina Blok, Rj Mical, Mo- hammad Norouzi, and Noah 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, Toronto, Canada, 2023. Association for Computational Lin- guistics. 3, 4, 5 [25] Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chun- rui Han, Guopeng Li, Yuang Peng, Quan Sun, Jingwei Wu, Yan Cai, Zheng Ge, Ranchen Ming, Lei Xia, Xianfang Zeng, Yibo Zhu, Binxing Jiao, Xiangyu Zhang, Gang Yu, and Daxin Jiang. Step1x-edit: A practical framework for gen- eral image editing, 2025. 3 [26] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 2 [27] Runnan Lu, Yuxuan Zhang, Jiaming Liu, Haofan Wang, and Yiren Song.Easytext: Controllable diffusion trans- former for multilingual text rendering.arXiv preprint arXiv:2505.24417, 2025. 3 [28] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Seamlessly ren- dering text with intricate spatial structures in text-to-image generation, 2023. 3, 4 [29] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose- guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 4117–4125, 2024. 2 [30] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. 3 [31] Yue Ma, Qingyan Bai, Hao Ouyang, Ka Leong Cheng, Qi- uyu Wang, Hongyu Liu, Zichen Liu, Haofan Wang, Jingye Chen, Yujun Shen, and Qifeng Chen. Calligrapher: Freestyle text image customization, 2025. 4 [32] Yue Ma, Kunyu Feng, Zhongyuan Hu, Xinyu Wang, Yucheng Wang, Mingzhe Zheng, Xuanhua He, Chenyang Zhu, Hongyu Liu, Yingqing He, et al. Controllable video generation: A survey.arXiv preprint arXiv:2507.16869, 2025. 2 [33] Yue Ma, Kunyu Feng, Xinhua Zhang, Hongyu Liu, David Junhao Zhang, Jinbo Xing, Yinhan Zhang, Ayden Yang, Zeyu Wang, and Qifeng Chen. Follow-your-creation: Empowering 4d creation through video inpainting. arXiv preprint arXiv:2506.04590, 2025. 2 [34] Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Leqi Shen, Chenyang Qi, Jixuan Ying, Chengfei Cai, Zhifeng Li, Heung-Yeung Shum, et al. Follow-your-click: Open-domain regional image animation via motion prompts. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 6018–6026, 2025. 3 [35] Yue Ma, Yulong Liu, Qiyuan Zhu, Ayden Yang, Kunyu Feng, Xinhua Zhang, Zhifeng Li, Sirui Han, Chenyang Qi, and Qifeng Chen. Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning. arXiv preprint arXiv:2506.05207, 2025. 2 [36] Yue Ma, Zexuan Yan, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, et al. Follow-your-emoji-faster: To- wards efficient, fine-controllable, and expressive freestyle portrait animation. arXiv preprint arXiv:2509.16630, 2025. 3 [37] Yue Ma, Zhikai Wang, Tianhao Ren, Mingzhe Zheng, Hongyu Liu, Jiayi Guo, Mark Fong, Yuxuan Xue, Zixi- ang Zhao, Konrad Schindler, et al.Fastvmt: Eliminat- ing redundancy in video motion transfer. arXiv preprint arXiv:2602.05551, 2026. 2 [38] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. 4 [39] OpenAI. Gpt-image-1, 2025. 4 [40] Nobuyuki Otsu. A threshold selection method from gray- level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62–66, 1979. 5 [41] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. 2, 3 [42] Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1–140:67, 2019. 4 11 [43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. 3 [44] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. S. Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text- to-image diffusion models with deep language understand- ing. ArXiv, abs/2205.11487, 2022. 4 [45] Yutao Shen, Junkun Yuan, Toru Aonishi, Hideki Nakayama, and Yue Ma. Follow-your-preference: Towards preference- aligned image inpainting. arXiv preprint arXiv:2509.23082, 2025. 3 [46] Wenda Shi, Yiren Song, Dengming Zhang, Jiaming Liu, and Xingxing Zou. Fonts: Text rendering with typography and style controls. arXiv preprint arXiv:2412.00136, 2024. [47] Wenda Shi, Yiren Song, Zihan Rao, Dengming Zhang, Ji- aming Liu, and Xingxing Zou. Wordcon: Word-level ty- pography control in scene text rendering. arXiv preprint arXiv:2506.21276, 2025. 3 [48] Ying Tai, Nikai Du, Rui Xie, Zhennan Chen, Qian Wang, Zhengkai Jiang, Kai Zhang, and Jian Yang. Investigating text insulation and attention mechanisms for complex visual text generation, 2026. 2, 3, 4, 5, 6 [49] Image Team, Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Xin Jin, Liangchen Li, Zhen Li, Zhong-Yu Li, David Liu, Dongyang Liu, Junhan Shi, Qilong Wu, Feng Yu, Chi Zhang, Shifeng Zhang, and Shilin Zhou. Z-image: An efficient im- age generation foundation model with single-stream diffu- sion transformer, 2025. 2, 3 [50] Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, et al. Kimi k2. 5: Visual agentic intelligence. arXiv preprint arXiv:2602.02276, 2026. 3, 7 [51] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text gener- ation and editing. In International Conference on Learning Representations, pages 56783–56799, 2024. 3, 5 [52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neu- ral Information Processing Systems. Curran Associates, Inc., 2017. 6 [53] Jiangshan Wang, Yue Ma, Jiayi Guo, Yicheng Xiao, Gao Huang, and Xiu Li. Cove: Unleashing the diffusion fea- ture correspondence for consistent video editing. Advances in Neural Information Processing Systems, 37:96541–96565, 2024. 3 [54] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Tam- ing rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. [55] Qinghe Wang, Xiaoyu Shi, Baolu Li, Weikang Bian, Quande Liu, Huchuan Lu, Xintao Wang, Pengfei Wan, Kun Gai, and Xu Jia. Multishotmaster: A controllable multi-shot video generation framework. arXiv preprint arXiv:2512.03041, 2025. [56] Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, De- qing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun Wen, Wensen Feng, Xiaoxiao Xu, Yi Wang, Yichang Zhang, Yongqiang Zhu, Yujia Wu, Yuxuan Cai, and Zenan Liu. Qwen-image technical report, 2025. 2, 3 [57] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raf- fel. Byt5: Towards a token-free future with pre-trained byte- to-byte models, 2022. 4 [58] Zexuan Yan, Yue Ma, Chang Zou, Wenteng Chen, Qifeng Chen, and Linfeng Zhang. Eedit: Rethinking the spatial and temporal redundancy for efficient image editing, 2025. 3 [59] Zhenyu Yan, Jian Wang, Aoqiang Wang, Yuhan Li, Wenxi- ang Shang, and Ran Lin. Textmaster: A unified framework for realistic text editing via glyph-style dual-control, 2025. 4 [60] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 4 [61] Ruijie Ye, Jiayi Zhang, Zhuoxin Liu, Zihao Zhu, Siyuan Yang, Li Li, Tianfu Fu, Franck Dernoncourt, Yue Zhao, Ji- acheng Zhu, Ryan Rossi, Wenhao Chai, and Zhengzhong Tu. Agent banana: High-fidelity image editing with agen- tic thinking and tooling, 2026. 4 [62] Jusheng Zhang, Kaitong Cai, Yijia Fan, Jian Wang, and Keze Wang. Cf-vlm:counterfactual vision-language fine-tuning, 2025. 3 [63] Jusheng Zhang, Yijia Fan, Wenjun Lin, Ruiqi Chen, Haoyi Jiang, Wenhao Chai, Jian Wang, and Keze Wang. GAM- agent: Game-theoretic and uncertainty-aware collaboration for complex visual reasoning.In The Thirty-ninth An- nual Conference on Neural Information Processing Systems, 2025. [64] Jusheng Zhang, Zimeng Huang, Yijia Fan, Ningyuan Liu, Mingyan Li, Zhuojie Yang, Jiawei Yao, Jian Wang, and Keze Wang.KABB: Knowledge-aware bayesian bandits for dynamic expert coordination in multi-agent systems. In Forty-second International Conference on Machine Learn- ing, 2025. 3 [65] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction- guided image editing. In Advances in Neural Information Processing Systems, 2023. 4 [66] Ruiqiang Zhang, Hengyi Wang, Chang Liu, Guanjie Wang, Zehua Ma, and Weiming Zhang. Freetext: Training-free text rendering in diffusion transformers via attention localization and spectral glyph injection, 2026. 2 [67] Zhilin Zhang, Xiang Zhang, Jiaqi Wei, Yiwei Xu, and Chenyu You. Postergen: Aesthetic-aware paper-to-poster generation via multi-agent llms. arXiv:2508.17188, 2025. 4 12 [68] Jun Zhou, Jiahao Li, Zunnan Xu, Hanhui Li, Yiji Cheng, Fa-Ting Hong, Qin Lin, Qinglin Lu, and Xiaodan Liang. Fireedit: Fine-grained instruction-based image editing via region-aware vision language model, 2025. 4 [69] Chenyang Zhu, Kai Li, Yue Ma, Longxiang Tang, Chengyu Fang, Chubin Chen, Qifeng Chen, and Xiu Li. Instantswap: Fast customized concept swapping across sharp shape differ- ences. arXiv preprint arXiv:2412.01197, 2024. 3 [70] Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems, 2025. 4 [71] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. pages 89–106, 2024. 4 [72] Yushen Zuo, Qi Zheng, Mingyang Wu, Xinrui Jiang, Ren- jie Li, Jian Wang, Yide Zhang, Gengchen Mai, Lihong V. Wang, James Zou, Xiaoyu Wang, Ming-Hsuan Yang, and Zhengzhong Tu. 4kagent: Agentic any image to 4k super- resolution. 2025. 4 13 A. More Qualitative Results This section provides additional qualitative examples as shown in Fig. 9, 10, 11 for visual comparison. 14 Original w/o Refine Iterative Refine Originalw/o Refine Iterative Refine Figure 9. More qualitative results for refinement process. 15 A documentary poster displays "Resplendent" in warm colors. A white background displaying \"|1 1 1 |\", \"|a b c |\", \"|a² b² c²|\" in mathematical notation. Qwen Image A theoretical physics paper presents \"F=-k_BTlnZ\"in complex notation. A classroom banner displays \"Star\"in twinkling letters A thermodynamics lecture shows \"λ_dB=h/p=h/(mv)\" on a green chalkboard A lecture board displaying \"Av=λv\", \"(A-λI)v=0\", \"det(A-λI)=0\"in clear script. 一块⻛化的招牌上用传统书法 写着\"霡霂（小雨）\" A classroom blackboard bears \"Gmm/r²\"in handwritten script. OursFluxdevFluxkleinGLMImage Figure 10. More qualitative results for GlyphBanana, Qwen-Image as base model. 16 A poster displaying \"iℏ∂ψ/∂t=Ĥψ\", \"Ĥ=-ℏ²/2m∇²+V\", \"Ĥψ=Eψ\" for educational purposes." 玩具盒上用可爱粉色字体写着\"娃\" "A science museum exhibit shows \"∇×(∇×F)=∇(∇·F)-∇²F\" on an information panel." An academic journal page displays \"∮_C Pdx+Qdy=∫_D (∂Q/∂x- ∂P/∂y)dσ\" in serif typography." An art installation features \"Mellifluous\" in metallic letters. A physics lecture board shows \"E²=(pc)²+(m₀c²)²\" written in white chalk. 一块风化的招牌上用传统书法写着\" 遮莫（不管，任凭）\" 一扇古旧的窗户上发黄的羊皮纸写着 \"泬寥（空旷而清朗的样子）\" Z-ImageOursGLM-ImageFLUX-devTex t C rafte r Figure 11. More qualitative results for GlyphBanana, Z-Image as base model. 17 Table 4. Illustration of GlyphBanana-Benchmark. The bench- mark contains multimodal inputs spanning English, Chinese, and scientific-formula subsets, together with reference images and masks, and follows a ladder-shaped difficulty design. Avg.|Text| denotes the average length of the target rendered text, and Avg.|Prompt| denotes the average length of the corresponding prompt. SubsetNum. Avg.|Text| Avg.|Prompt| English Subsets GlyphBanana-En (Easy)504.0847.74 GlyphBanana-En (Rare)258.9256.84 Chinese Subsets GlyphBanana-Zh (Easy)502.0019.00 GlyphBanana-Zh (Rare)2511.2027.48 Scientific Subsets (Ladder Difficulty) GlyphBanana-F (Easy)356.4662.46 GlyphBanana-F (Mid)4516.0472.64 GlyphBanana-F (Hard L1)4041.4194.12 GlyphBanana-F (Hard L2)20303.35376.85 Total / Average29032.6876.62 B. Benchmark Statistics B.1. GlyphBanana-Benchmark Overview Table 4 reports the detailed statistics of GlyphBanana- Benchmark. The benchmark spans English, Chinese, and scientific-formula subsets, and the formula branch follows a ladder-shaped difficulty schedule from short expressions to long multi-line structures. This progression is useful for stress-testing both the layout planner and the auxiliary ren- dering tools under increasingly complex text lengths and prompt conditions. C. Layout Planner Agent C.1. VLM-Based Text Grounding with Auxiliary Tools Table 5 summarizes the ablation study for the layout plan- ner, where IoU measures the overlap between predicted and ground-truth bounding boxes for formula placement. The VLM only setting uses the VLM without any coordinate- grid overlay as the baseline, while the remaining variants equip the same planner with grids of different densities. The Table 5. Ablation Study on VLM-Based Text Grounding with Auxiliary Tools. ConfigurationMean IoU↑ Median IoU↑ Std↓ Improvement VLM only)0.27030.25080.1620– VLM + 3×3 Grid0.44750.38920.1464+65.6% VLM + 5×5 Grid0.55310.54060.1280+104.6% VLM + 8×8 Grid0.37760.36280.1552+39.7% key conclusion is that moderate-density coordinate aids are most effective: adding a 5×5 grid raises mean IoU from 0.2703 to 0.5531, corresponding to a 104.6% improvement over the VLM-only baseline. In contrast, the 8×8 grid achieves only a 39.7% gain, suggesting that overly dense visual guides introduce clutter and weaken spatial ground- ing. This observation motivates the current planner design, which uses a coordinate overlay to improve spatial ground- ing while explicitly instructing the VLM to ignore the red guide lines when describing scene content. C.2. Formula Renderer as an Auxiliary Tool The formula renderer provides a deterministic auxiliary tool for synthesizing the glyph template used by the down- stream injection stage.According to the implementa- tion in infer/formula helper.py, the tool first de- tects whether the input should be treated as mathemati- cal content, converts Unicode math symbols into LaTeX- compatible expressions when needed, performs lightweight automatic line breaking for long expressions, and then dis- patches the content to a renderer selected by capability. For LaTeX-like content, the preferred route is MathJax through a Node.js backend; if that path is unavailable, the system falls back to matplotlib mathtext; otherwise, plain text is rendered with PIL and a font selected from the available registry. Multi-line expressions are rendered line by line and then vertically composed, after which the final glyph canvas can optionally be rotated to match the planned lay- out. C.3. Glyph Template Injection Illustration Once the auxiliary renderer produces a glyph template aligned with the typography plan, GlyphBanana injects that template into the latent-space refinement process.Fig- ure 13 visualizes this stage and complements the main- paper method description by showing how the rendered glyph prior is fused with the diffusion latent while preserv- ing the surrounding scene structure. 18 Input text + bbox (+ color / weight / font) Unicode-to-LaTeX conversion and auto line breaking LaTeX-like content? MathJax available? Render with MathJax SVG → PNG Fallback to matplotlib mathtext Render plain text with PIL + font registry Compose to target canvas, stack multi-line re- sults, apply rotation Glyph template image No Yes Yes No Figure 12. Execution flow of the formula-rendering auxiliary tool used by GlyphBanana. The implementation prefers MathJax for rich LaTeX formulas, falls back to matplotlib mathtext when Node.js or SVG conversion is unavailable, and uses PIL-based text rendering for non-LaTeX content. Early Stage Later Stage Glyph Latents Attention Enhancing 푥 ! ~σ 푥 " ~퐷 #$# Figure 13. Schematic diagram of enhanced text rendering by injecting glyph templates in latent space. D. VLM Agent Prompt Templates This section documents the prompt templates used by the current VLM agent implementation. The prompt stack is 19 designed to be model-agnostic and can be attached to dif- ferent diffusion backbones as long as they support the re- quired conditioning interfaces. In a representative work- flow, a text-to-image diffusion model first produces a ref- erence image, after which the VLM planner infers a struc- tured typography plan from the reference image, the user prompt, and the target text contents. The clean-prompt and style-prompt modules then support background regenera- tion, glyph injection, and subsequent harmonization with an image-to-image diffusion model. D.1. Typography Analysis Prompt Scenario. This prompt is invoked after Stage 2 reference- image generation and before any glyph injection is pre- pared. It is used only when the user does not manually override text regions. Function. Its role is to transform an unstructured visual reference into a machine-readable typography plan that specifies both global scene attributes and per-region render- ing instructions. Inputs and outputs. The call consumes four pieces of in- formation: the Stage 2 reference image, the original user prompt, the list of text or formula contents to be ren- dered, and a dynamically generated font list.The re- turned output is a strict JSON object with two top-level fields, image analysis and textregions. The for- mer provides scene-level descriptors such as background style, dominant colors, and text style hints; the latter pro- vides region-level attributes such as bounding boxes, font choice, color, alignment, and rotation. Dependencies. Thepromptdependsonthe grid-overlay utility,the font registry exposed by infer/formula helper.py, and the VLM backend configured in VLMAgent. The resulting typography plan is later consumed by both the glyph injector and the Stage 4 style harmonizer. Prompt: Typography Analysis You are an expert in image typography analysis. Given a reference image with a 5×5 grid and coordinate anno- tations, analyze the natural text rendering style and over- all scene. Then plan the best typography layout for each text/formula item. Critical constraints. • CRITICAL: The reference image shows text that is FLAT and FACING the screen directly (frontal view, no perspective distortion). The planned boxes must also remain flat and frontal, with parallel top and bot- tom edges and no angled or perspective-distorted re- gions. • IMPORTANT: The red grid lines and coordinate la- bels are ONLY positioning aids added by the sys- tem. They are not part of the actual image and must be ignored when describing background style, dominant colors, and textstylehint. • The5×5gridusesthecoordinates 0.0, 0.2, 0.4, 0.6, 0.8, 1.0oneachaxisand provides the positioning reference for normalized coordinates in [0, 1]. Per-region fields. • content: target text or formula. • bbox: [x min ,y min ,x max ,y max ] in [0, 1], flat and horizontal. • font: a font selected from the registered font list, or auto. • font weight: light/regular/bold. • font sizeratio: scalar in [0.1, 1.0] relative to box height. • color:one of white, black, red, blue, green, yellow, orange, brown, gray, gold, silver, purple, pink. • is latex: boolean flag indicating whether the re- gion is a formula. • alignment: left/center/right. • rotation: rotation angle in degrees, with 0 denot- ing horizontal text. Available fonts are provided dynamically through the placeholder font list.The output is re- quired to be strict JSON with two top-level entries, imageanalysis and textregions. D.2. Generate Clean Prompt Scenario. This prompt is called at the beginning of Stage 3, after the typography plan has already been pro- duced and immediately before background denoising and glyph injection. Function. Its role is to remove explicit text-rendering in- structions from the original prompt so that the diffusion backbone can focus on regenerating a clean background rather than hallucinating additional text. Inputs and outputs. The interface accepts the original prompt and optionally a typography plan. In the current VLMAgent implementation, the function signature still ex- poses typography plan, but the active call path only forwards the original prompt text into the VLM prompt body. The output is a single rewritten clean prompt string. Dependencies. This prompt depends on the original user prompt and the VLM backend. Its output is then fed di- rectly into the Stage 3 denoising step, where it conditions 20 the background generation used for subsequent pixel-space text compositing and latent injection. Prompt: Generate Clean Prompt Remove ALL quoted text, formulas, and text-rendering instructions from the prompt.Keep ONLY the scene/background/style description. Add “no text visi- ble” at the end. Examples. • Input: A classroom blackboard displays “E=mc²” in elegant chalk writing. Output: An empty classroom blackboard as back- ground, clear and without any text. No text visible. • Input: A stone monument is engraved with “Knowl- edge is power”, surrounded by a bamboo grove. Output: A blank stone monument as background, sur- rounded by a bamboo grove, clear and without any text. No text visible. Output ONLY the cleaned prompt, nothing else. D.3. Generate Style Prompt Scenario. This prompt is used in Stage 4 when an image- to-image diffusion model is employed for style harmo- nization. It is called after Stage 3 has produced the in- jected image and after the planner has already produced image analysis. Function. Its purpose is to compress the scene-level anal- ysis into a short editing instruction that preserves the back- ground while restyling the foreground text or formulas so that they better harmonize with the image. Inputs and outputs. The input is the imageanalysis field of the typography plan, specifically the background style, dominant colors, and text-style hint. The output is a short English editing prompt, typically 10–30 words, which is then forwarded to the image-to-image diffusion model used for harmonization. Dependencies. This prompt depends on the success of the typography-analysis stage, because it reuses the planner’s scene descriptors instead of reading the image again. Its downstream dependency is the Stage 4 image-to-image dif- fusion model, which consumes the resulting editing instruc- tion as its conditioning prompt. Prompt: Generate Style Prompt You generate a SHORT image-editing instruction (10–30 words) for a style-transfer model. Goal: restyle fore- ground text to harmonize with the background while keeping the background untouched. Do NOT move, re- size, or alter any text content or position. Examples. • Input: background style="weathered stone wall", colors=["#8B7D6B","#A09080"], hint="carved stone lettering" Output: Restyle text as deeply carved stone engrav- ings matching the weathered wall texture and earthy tones. • Input: background style="neon-lit cyberpunk street", colors=["#F00F","#00F"], hint="glowing neon sign" Output: Make text glow like neon signs with magenta and cyan edges against the dark street scene. • Input: background style="minimalist white paper", colors=["#F","#E0E0E0"], hint="clean printed type" Output: Render text as crisp black ink print on the clean white background with subtle shadow. Output ONLY the instruction, nothing else. It must be in English and 10–30 words long. D.4. Refine Prompt Scenario. This prompt is a reserved interface in VLMAgent for generic prompt enhancement. It is optional in the overall pipeline and may be enabled or disabled de- pending on the target diffusion backbone and deployment strategy. Function. Its goal is to rewrite a user prompt into a more rendering-friendly form while preserving the quoted text exactly. Inputs and outputs. The interface accepts the original prompt, an optional text-content hint, the number of vari- ants to sample, and a temperature value. It returns one or more rewritten prompt strings. Dependencies. The prompt depends only on the VLM backend. In some deployments, this refinement step may be replaced by deterministic prompt normalization, so the VLM-based refiner remains optional rather than mandatory. Prompt: Refine Prompt You are a prompt engineer for a text-to-image model that renders text inside images. 1. Keep the original scene description and ALL quoted 21 text exactly as-is. 2. Add that text should be clearly legible, well- positioned, and high-contrast. 3. Add brief visual details (lighting, style, materials) that make the scene vivid. Output only the enhanced prompt, nothing else. D.5. Score Image Prompt Scenario. This prompt defines a generic absolute image scorer in VLMAgent. It is not required by the core pipeline, because many deployments instead rely on OCR-based se- lection or external evaluation metrics. Function. Its purpose is to assign a single scalar score to one generated image by jointly considering image quality, prompt alignment, and text readability. Inputs and outputs. The interface takes one image to- gether with the corresponding prompt and an optional ex- plicit text-content string. It returns a single floating-point score in the range [0, 10]. Dependencies. This prompt depends on the VLM back- end and a parsed image input. It is kept as a reusable evalu- ation primitive for alternative pipelines, future ablations, or backbone-specific selection strategies. Prompt: Score Image Rate this image 0–10 based on: • overall quality (clarity, color, composition): 0–3; • alignment with prompt: 0–4; • text accuracy and readability (if applicable): 0–3. Output only the numeric score, nothing else. D.6. Rank Images Prompt Scenario. This prompt defines a generic multi-image ranking interface in VLMAgent. Similar to the single- image scorer, it is optional and can be switched on when a deployment prefers VLM-based ranking over OCR-based candidate selection. Function. Its role is to sort several candidate images from best to worst under shared criteria, so that rank positions can be converted into stepwise scores. Inputs and outputs. The call accepts a list of candidate images, the original prompt, and optionally the expected text string. It returns an ordered index list, which the im- plementation then maps to descending scores. Dependencies. The prompt depends on a multi-image VLM call. In many text-rendering settings, this function- ality is superseded by OCR-based selection, which is more directly tied to rendering precision. Prompt: Rank Images Rank thesen images from best to worst based on over- all quality, prompt alignment, and text accuracy (if ap- plicable). Output only the ranking as comma-separated indices (for example, 3,1,4,2), nothing else. E. Evaluation Interfaces This section summarizes the prompt-based evaluation in- terfaces associated with the VLM agent. In a representa- tive model-agnostic deployment, candidate selection can be performed by OCR-based scoring, which asks the VLM to transcribe the rendered text and compares the result against the target string extracted from the prompt. Other inter- faces, including style scoring, faithfulness scoring, VQAS- core, and CLIPScore, remain reusable evaluation compo- nents that can be enabled or disabled depending on the eval- uation protocol. E.1. API-Based OCR Recognition Scenario. This prompt is used for final candidate selec- tion after the reference image, injected result, and harmo- nized variants have all been generated. Function. Its role is to directly transcribe the visible text from each candidate image so that the system can compare recognized text against the target content and choose the most accurate rendering. Inputs and outputs. For the API-based VLM evaluator, the expected text T is extracted from the quoted spans in the input prompt. Each call then consumes one candidate image and returns a raw recognized text string with no explanation. The resulting transcription is subsequently compared with T using edit-distance metrics. Dependencies. This prompt depends on the image can- didates produced by the generation pipeline, the quote- based text extractor in eval/core/metrics.py, and the VLM backend.It is a natural selector for model- agnostic text-rendering systems because it directly mea- sures rendered-text fidelity rather than relying on backbone- specific confidence signals. The model is asked to directly transcribe the rendered text: 22 Prompt: OCR Recognition Please read and output ALL the text content visible in this image. Only output the text you can see, nothing else. If there are multiple text elements, separate them with spaces. Do not add any explanations or descriptions, just the raw text content. Let N(·) denote lowercase normalization with whites- pace collapsing, and let d Lev be the Levenshtein distance. If R is the recognized text, the VLM-based text scores are computed as Acc VLM = max 0, 1− d Lev (N(T),N(R)) |N(T)| ,(7) NED VLM = max 0, 1− d Lev (N(T),N(R)) max (|N(T)|,|N(R)|) + ε , (8) where ε is a small constant for numerical stability. For the standalone OCR metric used in the benchmark tables, we additionally report MinerU-based OCR scores with the same edit-distance formulation after normalizing the recognized text. E.2. VLM Style Score Scenario. This prompt belongs to the evaluation toolkit rather than the active generation path. Function. It estimates image-level style and quality com- patibility through a direct scalar judgment on a 0–10 scale. Inputs and outputs. The interface takes a single gener- ated image as input and returns one scalar score, which is normalized into S style . Dependencies. It depends only on the VLM evaluation backend and does not require the original prompt text. The VLM style score is implemented as a direct quality judgment on a 0–10 scale: Prompt: VLM Style Score Evaluate the overall quality of this image considering: 1. image clarity and sharpness; 2. visual coherence and aesthetics; 3. proper rendering of all elements. Rate from 0–10, respond with only a number. If the returned scalar is s style ∈ [0, 10], we normalize it as S style = s style 10 . E.3. VLM Faithfulness Score Scenario. This prompt is used in the evaluation module to quantify prompt adherence, but it is not invoked during the main generation loop. Function. Its goal is to measure whether the generated image remains faithful to the full prompt, including scene description, object presence, style intent, and text place- ment. Inputs and outputs. The interface consumes one gener- ated image together with the original prompt and returns one scalar score in the range [0, 10], which is then normal- ized into S faith . Dependencies. It depends on both the image and the orig- inal prompt text, because faithfulness is defined relative to the complete semantic condition rather than OCR accuracy alone. Prompt faithfulness is measured by asking the VLM to jointly assess scene consistency, object completeness, style fidelity, and text placement: Prompt: Faithfulness Score You are evaluating how faithfully this generated image matches its text prompt. Prompt: “original prompt” Consider the following aspects: 1. Scene & background: does the scene match the de- scription? 2. Objects & elements: are all described objects and el- ements present? 3. Style & color:does the visual style match the prompt’s intent? 4. Text content & placement: is the text rendered in the correct location with correct content? Rate the overall faithfulness from 0–10, respond with only a number. If the raw response is s faith ∈ [0, 10], we use the normal- ized score S faith = s faith 10 . E.4. VQAScore Interface Scenario. This interface is part of the evaluation stack and is independent of the generation-time prompt calls in the model-agnostic pipeline. Function. Its role is to compute a paired image-text rele- vance score without additional prompt engineering. 23 Inputs and outputs. The interface takes the generated image path and the original prompt P as its text query. The output is a scalar relevance score returned by the clip-flant5-xxl-based VQAScore model. Dependencies. It depends on the local VQAScore wrap- per under eval/TextCrafter Eval/vqascore.py and the underlying t2vmetrics implementation. For VQAScore, we do not perform extra prompt engineer- ing. Instead, the original prompt P is directly used as the text query paired with image I in the clip-flant5-xxl scorer: Text Query for VQAScore Input text to VQAScore: the original prompt P itself, without additional instructions or template wrapping. The resulting score is S VQA = f VQA (I,P),(9) where f VQA denotes the paired image-text score returned by the VQAScore model. E.5. CLIP Score Scenario. This metric is used only in the evaluation mod- ule and is not part of the runtime prompt workflow of the model-agnostic generation pipeline. Function. It measures global image-text alignment be- tween the generated result and the original prompt. Inputs and outputs. The implementation takes one im- age path and one prompt string, prepends the fixed text prefix “A photo depicts” to the prompt, and returns a non- negative scalar CLIPScore. Dependencies. It depends on the CLIP ViT-L/14 encoder loaded in eval/core/metrics.py. The image and text embeddings are normalized before their cosine simi- larity is rescaled into the final score. For image I and prompt P , CLIP produces image and text embeddings, denoted by φ img (I) and φ text (P). The im- plementation converts cosine similarity into a non-negative score via S CLIP = 2.5· max φ img (I) ⊤ φ text (P) ∥φ img (I)∥ 2 ∥φ text (P)∥ 2 , 0 .(10) 24