1 Introduction↩︎

Existing generative models have made remarkable progress in image generation and editing, demonstrating the capability to synthesize images with high aesthetic quality and strong alignment with text prompts. Concurrently, with the evolution of generative models, reasoning-informed generation has emerged as a prevailing trend and a highly active research direction. By unifying generation and understanding, this paradigm fosters mutual advancement alongside the development of unified multimodal models. Although several benchmarks have been established for reasoning-informed generation, they reveal that current generative models still exhibit limited performance in scenarios requiring precise comprehension, complex structural analysis, and deep reasoning. However, there remains a critical shortage of datasets dedicated to assisting and supporting this specific scenario.

3pt

To propel the development of reasoning-informed generation, we target the multidisciplinary domain and aim to propose a dataset tailored for reasoning-informed generation within this field. Compared to natural images, disciplinary illustrations prioritize the correctness of underlying structures and academic knowledge over sheer aesthetic appeal: a small mistake in a symbol, label, graph edge, molecular group, or notation can invalidate the result. Furthermore, the reasoning required here transcends the straightforward rewriting seen in natural image benchmarks such as WISE [6] and RISE [7], which predominantly rely on general world knowledge or commonsense concepts, where the prompts remain simple descriptions of the visual scene. Instead, disciplinary prompts are embedded with rich academic knowledge, thereby offering a rigorous testbed to elevate models’ image generation capabilities under complex reasoning. As summarized in Table 1, existing structured or text-rich visual datasets usually cover only part of this space, often lacking either multidisciplinary scope, joint T2I/editing supervision, or explicit reasoning and knowledge annotations. This motivates DisciplineGen-1M, which targets multidisciplinary academic visuals with both generation and editing formats under structured, knowledge-intensive supervision.

We address this gap with DisciplineGen-1M, which contains 1.2M samples for multidisciplinary text-to-image generation and image editing. To handle the heterogeneity of academic images, we combine four complementary construction pipelines: vector-graphics rendering, OCR-based editing, specialized programmatic synthesis, and large-scale T2I filtering. Together, these pipelines produce 1.2M examples across ten disciplines, covering both T2I and editing supervision for diagrams, charts, formulas, maps, molecular structures, music notation, board states, and other structured academic visuals.

Building on our dataset, we introduce a discipline-informed reasoning-generation model for multidisciplinary visual generation. We adapt the corresponding Qwen-Image generator for each task under a shared reasoning-generation pipeline. Qwen3-VL-8B [8] turns implicit disciplinary requests into explicit reasoning plans, while Qwen-Image-2512 handles text-to-image generation and Qwen-Image-Edit-2511 handles editing [9]. We first evaluate on discipline-related benchmarks: on GenExam [10], our T2I model reaches 51.4 relaxed score, and on GRADE [11], our editing model reaches 58.7 relaxed score, both ranking first among open-source models. We then evaluate on general reasoning-informed benchmarks, WISE [6] and RISE [7], observing consistent improvements beyond discipline-specific benchmarks. These results suggest that, beyond scale, exposing models to disciplinary structures, verifiable edits, and explicit knowledge constraints is important for academic visual generation and editing.

Our main contributions are summarized as follows:

1. We introduce a scalable multidisciplinary data construction framework that combines vector-graphics rendering, OCR-based editing, programmatic domain engines, and large-scale filtering.

2. We build DisciplineGen-1M, which to the best of our knowledge is the first million-scale multidisciplinary visual generation dataset.

3. We adapt open-source image generators with DisciplineGen-1M and demonstrate strong open-source state-of-the-art results on GenExam and GRADE, with additional generalization gains on WISE and RISE, validating the usefulness of DisciplineGen-1M for knowledge-grounded visual creation.

2 Related Work↩︎

2.1 Models for Image Generation and Editing↩︎

Recent text-to-image systems, largely driven by diffusion models [12], [13], have rapidly improved visual fidelity, language alignment, and controllability through models such as GLIDE [14], Imagen [15], latent diffusion models [16], Parti [17], and ControlNet [18]. Instruction-based editing has followed a similar trajectory, from attention-control and synthetic-instruction methods such as Prompt-to-Prompt [19] and InstructPix2Pix [20] to stronger editing systems such as SmartEdit [21], ICEdit [22], DreamOmni [23], and Qwen-Image [9]. These models are increasingly capable on natural images and design-like scenes, but their progress does not directly address the structured correctness requirements of academic visuals.

Recent unified multimodal systems further connect understanding and generation, as in Chameleon [24], Emu [25], OmniGen [26], Janus-Pro [27], BAGEL [28], and InternVL-U [29]. Such architectures can help parse complex instructions before synthesis, but stronger model capacity alone does not ensure correctness on disciplinary diagrams. DisciplineGen-1M complements these model advances with structured supervision that explicitly exposes models to academic image formats, symbolic constraints, and verifiable transformations.

2.2 Datasets for Image Generation and Editing↩︎

Large-scale image generation relies on broad image–text corpora such as Conceptual 12M [30], WIT [31], and LAION-5B [32], as well as prompt- or caption-enriched resources such as DiffusionDB [33], PixelProse [34], and TextAtlas5M [4]. These datasets provide scale and diversity, but their supervision is still dominated by web captions, natural-image aesthetics, and surface-level alignment. Instruction-editing datasets, including InstructPix2Pix [20], MagicBrush [35], UltraEdit [36], OmniEdit [37], AnyEdit [38], Pico-Banana-400K [39], and ScaleEdit-12M [40], similarly improve general editing behavior but remain focused on natural images, artworks, and object-level changes.

Recent specialized resources move closer to our setting: TextAtlas5M [4] emphasizes text-rich generation, and StructVisuals [3] studies structured charts and diagrams. Meanwhile, discipline-related benchmarks such as GenExam [10] and GRADE [11] show through evaluation that current models struggle with discipline-informed generation and editing. General reasoning-informed benchmarks such as WISE [6] and RISE [7] further evaluate world-knowledge-informed generation and reasoning-informed editing in broader settings. However, these efforts do not provide a large-scale training resource spanning both T2I and editing across many academic domains. DisciplineGen-1M fills this gap with multidisciplinary, structured, and verifiable supervision.

3 Data Construction↩︎

To obtain a large-scale multidisciplinary text-to-image (T2I) and image editing dataset, we design the construction framework shown in Figure 1. The central challenge is that academic images are not a single visual domain. Some are symbolic vector diagrams, some are text-label completion tasks, some require domain-specific engines, and others can only be obtained by filtering noisy open-source educational images. We therefore combine four pipelines: (1) vector-graphics data rendering; (2) OCR-based editing; (3) large-scale T2I data filtering; and (4) specialized programmatic synthesis. Each pipeline emits a unified annotation format so that the resulting data can support both generation and editing training.

Figure 2 illustrates the resulting data format and visual diversity. A sample is organized by discipline and fine-grained subdomain, and then instantiated as either a T2I example with an image and caption-like prompt or an editing example with an input image, a target image, and a natural-language instruction.

3.1 Vector Graphics Data Rendering↩︎

Collecting multidisciplinary images from existing sources often leads to limited coverage, quality, and resolution. The problem is harder for editing because high-quality paired images are rarely available. Directly using a proprietary image editing model such as Nano Banana Pro [41] to create paired data is also expensive and may introduce uncontrolled semantic changes. Many academic diagrams, however, are naturally described by vector primitives: text labels, arrows, curves, axes, equations, symbols, and spatial relations. Motivated by recent progress in SVG and TikZ understanding and generation [42]–[45], we synthesize a large portion of DisciplineGen-1M through structured vector rendering. Operating on SVG/TikZ code gives us high-resolution images, deterministic re-rendering, and explicit control over semantic edits.

3.1.1 SVG-Based Data Rendering↩︎

As illustrated in Figure 1, the SVG branch constructs T2I and editing data for domains such as biology, physics, and economics. Each image is first represented as a structured semantic instance containing domain-specific fields, such as textual nodes, directed relations, curve parameters, reaction components, axis labels, and scientific constraints. The instance is mapped to a diagram template and rendered into SVG through sampled layout, style, and typography. The same structured representation also generates captions, which gives aligned image-caption pairs for T2I training. For editing data, we apply verified transformations to the structured representation and re-render both source and target images. Examples include removing or restoring labels, shifting curves, changing key annotations, or modifying reaction components.

The SVG branch also supports web image initialization: after extracting a candidate image and its context, Gemini-3-Flash [46] generates SVG code for both the original and edited diagrams, followed by automatic rendering and filtering. Compared with directly prompting Nano Banana Pro [41] to generate the edited output, this SVG-based route reduces the estimated cost from $0.16 to $0.03 per sample while preserving stronger geometric consistency.

3.1.2 TikZ-Based Data Rendering↩︎

The TikZ branch targets biology, geography, physics, chemistry, computer science, and economics, where many diagrams can be expressed as compilable graphics programs. We first build a seed library of 1,346 items covering 17 fine-grained knowledge points across six disciplines. Each seed is expanded into a drawing blueprint and translated into TikZ source. A compiler-in-the-loop repair process handles LaTeX errors, and a VLM judge filters rendered images according to correctness, clarity, and completeness. For editing, samples that contain explicit labels are rewritten by replacing each label with numbered placeholders and recompiling the same source file. This creates fill-in editing pairs that are pixel-aligned except for the annotation tokens.

3.2 OCR Editing Pipeline↩︎

Many educational figures contain informative labels whose removal and recovery naturally define editing tasks. The OCR editing pipeline turns such figures into paired data without asking an image model to hallucinate a target. First, all text instances and bounding boxes are extracted using PP-OCR v5 [47]. Second, the image and OCR results are passed to a VLM that selects regions suitable for masking, rejecting boxes that are too small, visually ambiguous, or semantically unimportant. Finally, selected boxes are masked to form the input image, while the original image serves as the target. The VLM also writes the editing instruction, producing aligned triplets of input image, instruction, and ground-truth image.

3.3 T2I Data Filtering Pipeline↩︎

Programmatic synthesis offers precise control but cannot cover every real-world academic visual style. We therefore filter large open-source educational image pools for T2I training. The pipeline first applies pHash [58] deduplication, benchmark-overlap removal, and resolution filtering. It then performs OCR[47] pre-screening to remove language-dependent or text-dominant images. Remaining images are scored by Qwen3-VL-8B-Instruct [8] along multiple dimensions, including discipline relevance, diagrammatic nature, visual complexity, text richness, and required domain knowledge. Rule-based post-filtering keeps non-photographic, structurally informative, moderately complex images that are suitable for discipline-specific T2I generation. Detailed thresholds and score distributions are provided in Appendix 6.

3.4 Specialized Synthesis Pipelines↩︎

Editing data requires paired images with high consistency, which are difficult to directly obtain from open-source datasets and web images. In addition, existing T2I images fail to cover certain domains with sufficient data. Therefore, we design diverse specialized synthesis pipelines with carefully curated data engines to synthesize editing and T2I data for a number of disciplinary domains. Table 2 summarizes the main branches.

3.5 Dataset Analysis↩︎

Figure 3 summarizes the current data distribution. Prompts are generally long and information-dense, with a median length of 73 words, reflecting the fact that many samples encode not only visual appearance but also disciplinary constraints. Images are also diverse in category, resolution, and aspect ratio. The distribution contains square-ish diagrams, landscape charts and maps, tall notation-like images, and wide structural visuals, which helps expose models to realistic academic layout variation. At the category level, charts/graphs, geometry, tables, maps, molecular structures, music notation, timelines, and biological anatomy provide complementary visual grammars rather than a single dominant natural-image style.

4 Experiment↩︎

4pt

4.1 Training Setup↩︎

Based on DisciplineGen-1M, we adapt the image generation backbone for each task setting under a shared two-stage reasoning-generation design. Given an implicit disciplinary instruction, Qwen3-VL-8B [8] first produces an explicit reasoning plan that identifies the relevant discipline knowledge, target visual elements, and required structural changes. The task-specific image generator then synthesizes the final output: Qwen-Image-2512 for T2I generation and Qwen-Image-Edit-2511 for editing [9]. We use LoRA [68] in each setting to preserve the base model’s general visual prior while adapting it to academic diagrams. Both reported generators are optimized with the same DiffSynth training pipeline using FlowMatch SFT loss, dynamic-resolution image loading, and an effective batch size of 8. Detailed hyperparameters are provided in Appendix 7.

For T2I, we train on the text-to-image portion of DisciplineGen-1M. For editing, we train on the editing portion of DisciplineGen-1M. We organize evaluation into two groups. First, we evaluate discipline-related benchmarks, using GenExam [10] for T2I and GRADE [11] for editing. Second, we evaluate general reasoning-informed benchmarks, using WISE_Verified [6] for world-knowledge-informed T2I generation and RISEBench [7] for temporal, causal, spatial, and logical reasoning in image editing. We follow the official evaluation protocol of each benchmark. Training details are provided in Appendix 6.

4.2 Discipline-Related Benchmarks↩︎

Table 3 reports GenExam results for discipline-informed T2I generation. Our T2I model achieves state-of-the-art performance among all open-source models with an overall score of 51.4, outperforming Qwen-Image-2512 by 15.6 points and FLUX.2 dev by 8.5 points. This result substantially reduces the gap between open-source and closed-source models, while leading closed-source systems still retain a clear advantage. The improvement is consistent across all ten disciplines, with especially large gains in chemistry, economics, music, computer science, and history. These results indicate that training on DisciplineGen-1M improves not only visual quality but also the ability to follow discipline-specific constraints such as molecular connectivity, graph structure, musical notation, and chart semantics. The remaining gap to the strongest closed-source models shows that multidisciplinary academic generation is still far from saturated.

Table 4 further breaks down GenExam performance by image type. Our model improves over Qwen-Image-2512 across all eight visual categories, with the largest gains on Other (\(+46.7\)), plots and charts (\(+21.7\)), chemical structures (\(+16.7\)), and diagrams (\(+16.3\)). It also outperforms the open-source baselines on six of the eight categories, indicating that DisciplineGen-1M is especially effective for structured visuals that require precise symbolic elements, layout relations, and domain-specific graphical conventions. The remaining weaknesses are maps and sheet music, where FLUX.2 dev still obtains higher scores, suggesting that geographic layout reasoning and dense music notation may require more targeted data or specialized rendering constraints. Compared with closed-source models, the gap remains substantial on highly structured formats such as maps, sheet music, and trees/graphs, highlighting these image types as useful directions for future data expansion.

Table 5 reports GRADE results for discipline-informed image editing. Our editing model reaches 58.7 overall, which is 14.4 points higher than the best listed open-source baseline, DreamOmni, and 28.7 points higher than Qwen-edit-2511. The gains are broad across physics, chemistry, mathematics, geography, biology, computer science, and history, suggesting that paired structured data is especially helpful for edits that must preserve most of the input while changing a small number of semantically important elements. Compared with closed-source models, it still lags behind GPT-Image-2 and Nano Banana Pro, but substantially narrows the open-source gap in discipline-informed editing.

4.3 General Reasoning-Informed Benchmarks↩︎

Table 6 evaluates whether the T2I model generalizes beyond discipline-related benchmarks to WISE_Verified, a world-knowledge-informed generation benchmark. Our model obtains a 0.66 overall score, outperforming BAGEL (w/ CoT) by 0.03, FLUX.2-dev by 0.10, and Qwen-Image-2512 by 0.16. The gains are strongest on spatial, biology, physics, and chemistry categories, which is consistent with the structured scientific supervision in DisciplineGen-1M. At the same time, BAGEL remains stronger on culture and time, suggesting that disciplinary visual data transfers best to knowledge settings with explicit spatial, scientific, or structural constraints, while broader cultural commonsense may require complementary training sources.

Table 7 reports results on RISEBench, a general reasoning-informed editing benchmark covering temporal, causal, spatial, and logical reasoning. Our editing model achieves 26.4 overall, outperforming Qwen-Image-Edit-2511 by 7.0 points and BAGEL (thinking) by 10.8 points among open-source models. It also exceeds the strongest official closed-source baseline in overall score, mainly due to large gains on temporal, causal, and spatial reasoning. The logical category remains comparatively weaker than top closed-source systems, indicating that some abstract or symbolic reasoning edits still require stronger planning or more targeted paired data. Overall, the RISE results suggest that training on structured academic editing pairs improves not only discipline-specific editing in GRADE, but also broader reasoning-informed editing behavior.

4.4 Qualitative Results↩︎

Figure 4 presents qualitative text-to-image comparisons on representative GenExam prompts. Across biology, physics, chemistry, and computer science examples, our model better preserves the key structural constraints required by the instruction, such as part-level labels in the neuron diagram, the relative directions of current and electron motion, the axial/equatorial geometry of the molecule, and the ordered states and transitions of the DFA. In contrast, baseline models often produce visually plausible images but miss important labels, introduce irrelevant details, or violate the requested spatial and symbolic relations. These examples are consistent with the quantitative results, suggesting that training on DisciplineGen-1M improves both visual fidelity and discipline-specific correctness.

Figure 5 further shows qualitative editing comparisons on representative GRADE instructions. The examples require targeted modifications while preserving the surrounding academic structure, such as editing labels, symbols, graphical relations, or domain-specific components. Compared with baselines, our model more reliably performs the requested edit without disrupting unrelated visual elements, which supports the GRADE results in Table 5 and highlights the value of paired structured supervision for discipline-informed image editing.

4.5 Discussion↩︎

The results support two conclusions. First, the discipline-related results on GenExam and GRADE show that academic visual generation and editing benefit from training data that is both structured and multidisciplinary. Improvements are not limited to the domains with the largest apparent representation in the statistics, which suggests transfer across shared visual primitives such as labels, arrows, plots, graphs, and symbolic notation. Second, the general reasoning-informed results on WISE and RISE indicate that this transfer extends beyond discipline-specific evaluation. WISE shows stronger world-knowledge generation when the prompt involves spatial or scientific constraints, while RISE shows that paired supervision improves reasoning-informed editing, especially for temporal, causal, and spatial edits.

5 Conclusion↩︎

We present DisciplineGen-1M, a million-scale dataset that jointly supports text-to-image generation and image editing for structured multidisciplinary academic visuals. The dataset is built with four complementary pipelines that combine structured vector rendering, OCR-based editing, programmatic domain synthesis, and large-scale filtering. This design produces both T2I and editing data with rich disciplinary coverage and explicit structural supervision. Adapting open-source generators with DisciplineGen-1M yields state-of-the-art open-source performance on GenExam and GRADE, as well as strong generalization results on WISE and RISE. These results show that scaling academic visual data with verifiable structure is a promising path toward image generation models that are not only visually plausible but also knowledge-correct.

6 Dataset Construction Details↩︎

6.1 Sub-Discipline Distribution↩︎

Figure 6 shows the sub-discipline composition of DisciplineGen-1M, covering 1.2M samples in total. The distribution is intentionally broad but not uniform. Mathematics and chemistry form the two largest discipline groups, accounting for 33.3% and 26.6% of the data, respectively, because these domains provide many programmatically verifiable structures such as plane geometry, function plots, solid geometry, molecular structures, and reaction equations. Music, computer science, and physics contribute another substantial portion through sheet music, trees and graphs, finite-state machines, mechanics, circuits, and optics. The long tail covers sports, biology, history, geography, and economics, which increases visual and semantic diversity beyond the dominant scientific diagram types. At the sub-discipline level, large categories such as plane geometry, molecular structures, and sheet music are complemented by smaller but distinctive formats including maps, timelines, nutrition charts, finite-state machines, ecology diagrams, and macroeconomics plots. This mixture supports both scale and breadth: high-frequency categories provide enough examples for robust structural learning, while low-frequency categories expose models to specialized academic visual conventions.

6.2 SVG-Based Data Rendering↩︎

Rendering from Structured Semantic Instance. SVG-Based Data Rendering provides a structured-knowledge-driven pipeline for constructing discipline-specific image generation and image editing data in economics and biology. As input sources, representative knowledge points and typical diagrammatic content are first manually selected from openly licensed textbooks and educational resources [73]–[78]. In the transformation process, GPT-5 [79] is used as a structuring assistant to convert the selected content into machine-readable knowledge units, including fields such as topic, objects, textual elements, relational information, diagram type, coordinate semantics, and visual constraints. The system then invokes predefined SVG templates, parameter samplers, and rendering rules to automatically generate disciplinary images such as tables, flowcharts, classification trees, quadrant diagrams, and coordinate-curve plots, together with corresponding captions. For quality control, the structured knowledge units and rendered images are manually checked to ensure disciplinary correctness, visual clarity, and consistency with the source knowledge. The final outputs include image-caption pairs for image generation tasks and triplet samples for image editing tasks, consisting of the original image, editing instruction, and edited image. This process decouples disciplinary knowledge modeling from visual rendering, improving the controllability, interpretability, and subject consistency of the constructed data.

The SVG-based framework is instantiated for six diagram families: biological lifecycle diagrams, economics curve charts, biological curve charts, biochemical reaction equations, food web diagrams, and food chain diagrams. For lifecycle, food chain, and food web diagrams, the main visual elements are English text nodes connected by arrows or relations, rather than natural-image objects. For reaction equations, editable elements are textual reactants or products. For curve charts, editable elements include curve parameters, axes, key points, annotations, and domain-specific semantic states. Therefore, all diagrams are treated as structured scientific visualizations composed of graphical primitives and textual elements.

Rendering from Web Images. For web images, we use PaddleOCR [47] to extract images and their associated textual context. We then apply Qwen3-VL-8B [8] to remove samples from undesirable domains, such as illustrative textbook figures, real-world photographs in exam questions, mathematical formulas, and visually complex images that are difficult to convert into SVG format.

Next, each image and its context are provided to Gemini-3-Flash [46] to produce the SVG code of the original image. The model further determines whether the original image should be treated as the input or the output image, and generates both an editing prompt and the SVG code of the paired image. The two SVG files are then rendered into an image pair. The prompts generated by Gemini-3-Flash [46] are detailed and explicit, and can serve as step-by-step reasoning prompts. To evaluate discipline-specific knowledge and competence, we further extract and summarize the key information in these prompts to obtain simplified, implicit prompts as the final data.

Finally, we perform an additional filtering step on each triplet of <input_image, prompt, output_image> to assess data correctness and quality. Gemini-3-Flash [46] is used to remove samples containing rich text formatting, physical structural errors, or annotation errors.

This construction strategy provides several practical advantages. First, it reduces the cost of data collection and annotation by automatically generating images, captions, metadata, and editing instructions from reusable structured representations. Second, it improves controllability since both visual appearance and semantic content can be explicitly specified and sampled. Third, it enables fine-grained and verifiable editing, as each edit corresponds to a structured transformation rather than an uncontrolled pixel-level modification. Finally, it supports scalable dataset expansion: new diagram types, templates, scientific constraints, or editing operations can be incorporated without redesigning the entire data generation pipeline.

6.3 TikZ-Based Data Rendering↩︎

6.3.1 Method↩︎

We construct text-to-image generation and image editing tasks across Biology, Geography, Physics, Chemistry, Computer Science, and Economics, with visual representations spanning anatomical and structural diagrams, flowcharts, plotted curves, network and pedigree figures, geological cross-sections, directional maps, and ray-optics / circuit schematics. All pipelines draw from a shared LLM-generated seed library of 1346 items covering 17 fine-grained knowledge points across the six disciplines (e.g.cell-structure “fill-in”, “ray optics”, “supply-and-demand curves”); each item carries common fields (discipline, category, difficulty, knowledge-point name) plus per-KP custom fields that are flattened into a single topic string for downstream rendering.

(1) TikZ-Compiled Generation. Each seed is rewritten by the LLM into a drawing blueprint with discipline-specific LaTeX package hints, then translated into TikZ source. Inspired by the tool-use formulation of [44], pdflatex is exposed as an external tool the model iterates against: compile errors are fed back to repair the source (inner self-repair loop). Successful renders are scored by a vision-model judge along correctness, clarity, and completeness; failures send feedback and the previous PNG back to code generation (outer visual-retry loop). The description and final PNG form a text-to-image pair.

(2) TikZ Label Fill-in Editing. For seeds whose main diagram has passed scoring and carries a labels list (e.g.component names on a cell-structure figure), we apply a deterministic regex rewrite to the final .tex—replacing each label in place with numbered placeholders in first-occurrence order—and recompile to obtain a geometrically aligned blanked PNG and an answer key. Because both diagrams share one .tex source, they are pixel-identical except for the annotation tokens, forming a natural image-editing pair usable in both directions (annotated \(\leftrightarrow\) blanked) without alignment checks.

Seed Library Construction. The 17 knowledge-point expansion prompts are hand-authored once per KP; the LLM is then driven through multi-pass continuation under a per-pass batch budget to bypass gateway-side truncation on long responses, with names already produced relayed back so the model proposes only new items, and deduplication performed on the merged result. The resulting per-item record carries common fields and a handful of KP-specific custom fields such as labels, key_points, view, and curve_type, which the downstream loader concatenates with the item name into a single topic string.

Two-Track Prompt Families. The TikZ and image-model pipelines maintain strictly disjoint per-discipline prompt packages: the TikZ side ships LaTeX macro-package hints (e.g. circuitikz, chemfig, chronosys) while the image side ships pure art-direction guidance with all LaTeX and coordinate vocabulary explicitly forbidden (e.g.for Biology: “mitochondria: bean-shaped/ellipsoidal, double membrane, cristae clearly visible”). Cross-contamination at the description stage is avoided by template-level isolation.

Reflection-Style Feedback Loops. We observe that one-shot generations rarely pass quality control in either pipeline, but explicitly relaying the compiler’s exact errors or the judge’s pinpointed issues and corrective suggestions back to the generator substantially improves subsequent attempts, consistent with iterative self-refinement patterns observed in recent LLM systems. Both the inner compile-repair loop in (1) and the outer visual-retry loop shared by (1) and (3) are instances of this pattern.

Vision-Judge Rubric. The judge returns JSON containing three sub-scores (correctness, clarity, completeness) and an aggregate, plus two mandatory list fields knowledge_errors and overlap_issues that act as hard vetoes: a non-empty entry in either fails the diagram regardless of the aggregate. The regeneration feedback prepends veto items ahead of the free-form summary string to guarantee they cannot be silently dropped by the judge’s natural-language paraphrase.

Retries and Parallelism. The TikZ pipeline uses nested retries (inner compile/repair \(\times\) outer visual regeneration); the image pipeline retains only the outer loop. Both limits are configurable. Batch execution uses a thread-pool across topics, with per-topic exceptions absorbed into failure records so a single bad topic does not abort the run. Upstream seed expansion is constrained by API-gateway truncation on long responses and concurrency throttling, which forced a fallback to two parallel workers and multi-pass continuation with a per-response batch-size hint.

Trace and Reproducibility. Every LLM prompt, raw response, extracted TikZ source, compile log, and intermediate PDF/PNG is written to disk under a monotonic-counter filename encoding step, visual round, and compile attempt; a chronological directory listing replays the full pipeline run. A per-topic JSON metadata record stores topic, description, scores, retry counts, failure reason, and fill-in outcome, allowing successful topics to be skipped on resume while partial runs are wiped clean to keep the trace counter monotonic.

Dataset Export. Successful topics are aggregated into a JSONL dataset with relative image paths and automatic sharding once a single shard would exceed 50,000 lines. The natural-language description from step 1 is written into a chain-of-thought field for downstream training, and fill-in items additionally ship their answer keys. PNG dimensions are read directly from the IHDR chunk so the entire stack remains free of third-party Python dependencies.

6.3.2 Prompt↩︎

This section reproduces the full prompts for the LLM stages of the TikZ-compiled generation track (track (1) in the Method). That track drives the LLM through five staged prompts: a description-expansion stage, a code-generation stage, a compile-repair stage, and a two-step visual-quality loop (a vision judge plus a feedback-driven regeneration). The stage numbers below follow the pipeline’s internal labels; Stage 3 is the pdflatex compilation step itself, which issues no LLM prompt and is therefore omitted here. All prompts are issued in English. Braced tokens such as {topic}, {description}, {error_log}, {code}, {prev_code}, {feedback} and {threshold} are filled at run time; {subject_hints} is replaced by the per-discipline package list reproduced at the end of this section.

6.3.2.1 Stage 1 — Description Expansion.

Turns a knowledge point into a concrete, TikZ-shaped drawing description.

6.3.2.2 Stage 2 — TikZ Code Generation.

Turns the drawing description into a complete, compilable standalone document.

6.3.2.3 Stage 4 — Compile Repair.

Fixes code that failed to compile, from a filtered pdflatex error log. (Retried up to a configurable number of times before the topic is abandoned.)

6.3.2.4 Stage 5a — Visual Judge.

A vision model scores the rendered PNG and returns a strict JSON verdict; {threshold} is the passing score.

6.3.2.5 Stage 5b — Visual Regeneration.

When the judge rejects, the model is shown the previous render plus the judge feedback and asked to localize and fix. (Up to a configurable number of rounds.)

6.3.2.6 Per-Discipline Package Hints (subject_hints).

The same {subject_hints} block is injected into Stages 1, 2, 4 and 5b. It gives the model a curated list of available TikZ packages and drawing conventions, which sharply reduces failures where the model invents a non-existent package or hand-draws something a library already provides. One block is shown per discipline.

6.4 T2I Data Filtering Pipeline↩︎

We collect open-source datasets, including K12 [80], MV-MATH [81], MAVIS [82], and ChemEval [83], comprising 2.5M samples in total, to construct an automated filtering pipeline for text-to-image data. First, we perform global pHash [58] deduplication on the original images and remove images that are duplicate or highly similar to those in the target benchmarks (GenExam [10], GRADE [11], and RISE [7]), to ensure the independence of the evaluation sets. We then discard images with small sizes (short size < 512 pixels) to ensure that the retained samples contain sufficient visual information.

After the basic cleaning stage, we further apply an automated discipline-image filtering pipeline designed to retain visually informative, discipline-relevant educational diagrams while excluding text-dominant or natural-image samples. The pipeline operates in three stages.

OCR Pre-Screening. In the first stage, we perform lightweight OCR-based pre-screening before invoking the multimodal model. For each image, the OCR [47] module estimates the total number of detected characters, the number of text boxes, and the text-area ratio, i.e., the proportion of image area occupied by text. Images are directly rejected at this stage if the prompt contains Chinese characters, if the detected text count exceeds 280 characters, or if text occupies more than 13% of the image area, since such samples are likely to be language-dependent or overly text-centric rather than suitable for visual generation.

Multimodal Quality Assessment. For the remaining images, we use Qwen3-VL-8B-Instruct [8] to conduct structured multi-dimensional quality assessment. The model outputs judgments for whether the image is a real photograph (is_real_photo: 0/1), its suitability as a discipline-specific educational image (discipline_image_score: 0–10), its dependence on textual content (text_richness_score: 0–10), whether it belongs to an abstract/synthetic diagram domain rather than a natural-image domain (image_domain_score: 0/1), its visual complexity (image_complexity_score: 0–10), and the amount of domain knowledge required to understand it (discipline_knowledge_score: 0–10). In addition, the model predicts a knowledge-difficulty level, a coarse image type, a fine-grained image subtype, a discipline label, and a free-text rationale. The coarse image types are defined over a closed set including diagram, chart_plot, geometry, formula, map, molecule, circuit, screenshot, photo, mixed, and other, while the fine-grained taxonomy further distinguishes more than 90 educational subcategories, such as function graphs, geometric transformations, force analysis diagrams, molecular structures, optical paths, algorithm flowcharts, and geological maps. Table 8 summarizes the score distributions of the four main evaluation dimensions used in downstream filtering, namely discipline_image_score, text_richness_score, image_complexity_score, and discipline_knowledge_score.

Rule-Based Filtering. We then apply rule-based post-filtering that combines OCR statistics with model predictions. An image is rejected if it is classified as a real photograph, if its discipline-image score is below 4, if its text-richness score is below 10, if its image-domain score is not exactly 1, if its complexity score is below 4, or if its discipline-knowledge score is below 4. We further restrict the accepted set to a whitelist of educational visual formats, namely diagram, chart_plot, geometry, formula, map, molecule, and circuit, while excluding screenshots, photographs, mixed-format images, and uncategorized cases. This design intentionally imposes exact constraints on text richness and image domain, so that the final dataset consistently contains text-free, non-photographic, structurally clear, visually reproducible, and academically meaningful discipline images.

Dataset Construction. After filtering, the retained samples are further grouped by discipline categories such as mathematics, physics, chemistry, biology, geography, computer science, history, music, and sports for downstream discipline-specific text-to-image generation.

6.4.1 Prompt of Multimodal Quality Assessment↩︎

This section reproduces the prompt used for Multimodal Quality Assessment in the T2I Data Filtering stage. The model uses the image as the primary signal and the paired instruction text as supporting context, then returns a strict JSON record.

6.5 Specialized Synthesis Pipelines↩︎

6.5.1 Mathematics Data Synthesis↩︎

We organize the mathematics data construction into three major categories: functions, plane geometry, and solid geometry.

(1) Functions. Data are synthesized using the Matplotlib library to render a diverse set of function plots. The pipeline covers five task types: ) Symmetry, where plots of functions are displayed under odd/even transformations; ) Zeros and Poles, where the function’s roots and poles are highlighted; ) Tangent Lines, where the tangent at a specified point is drawn; ) Derivatives, where the plot of the derivative is rendered alongside the original function; ) Function Properties, where parameters of trigonometric functions such as amplitude, period, and phase are varied to generate diverse behaviors.

(2) Plane Geometry. Plane geometry data are generated and rendered using Matplotlib and MathCanvas [48], producing diverse planar figures for downstream editing tasks. The pipeline covers three task types: ) Points and Lines, where constructions such as perpendiculars or parallels through a point outside a line are drawn; ) Geometric Transformations, including rotation, translation, and reflection of planar figures; and ) Geometric Properties, such as medians and altitudes of triangles, tangents to circles, and angle bisectors within plane shapes.

(3) Solid Geometry. We generate solid geometry data using GeoGebra [49] and Matplotlib to construct a variety of 3D editing tasks. The pipeline covers four types of operations: ) Solid Symmetry, where a solid is mirrored either across a plane or with respect to a point; ) Solids of Revolution, where a shape is rotated around an axis; ) Solid Translation, where a solid is shifted along a vector; and ) Solid Projection, where the orthogonal projection of a solid onto the \(x\)-\(y\) plane is rendered.

6.5.2 History Data Synthesis↩︎

We construct two types of editing tasks for the history domain.

(1) Timeline Completion. We collect historical event entries from Wikimedia Commons and retrieve their associated images through the static resource. To ensure that each sample is grounded in structured historical metadata, we only retain event entries that are linked to at least one Wikimedia Commons image. The resulting event pool contains approximately 500 historical events spanning from 3000 BCE to 1900 CE.

During data generation, we randomly sample a batch of events from the pool and sort them chronologically to construct a timeline. The timeline is rendered with Matplotlib, where visual attributes such as colors, layouts, fonts, and line styles are randomly configured to improve visual diversity. For each rendered timeline, one event is randomly masked to produce the input image, while the complete timeline is used as the ground-truth image. This process yields 10K paired editing samples for the timeline completion task.

(2) Historical Map Completion. The input and ground-truth map images are rendered using a Python-based geospatial visualization pipeline, primarily built upon Matplotlib and GeoPandas for rasterizing GeoJSON vector data. To improve visual diversity, we adopt multiple theme-specific color palettes and randomly sample rendering styles, producing maps with varied background colors, region colors, label styles, and visual layouts. Given the input GeoJSON data, the pipeline first performs random spatial cropping to generate local map regions. To control the difficulty of each sample, we further apply an adaptive scale search strategy, where the cropping ratio is iteratively adjusted through binary search until the number of visible labels falls within a predefined range, i.e., 5-8 labels. This ensures that each generated map contains sufficient textual content while avoiding overly crowded layouts.

To improve label readability, we introduce an iterative label de-overlapping algorithm. Specifically, the pipeline computes the spatial overlap between text bounding boxes and applies repulsive offsets to conflicting labels, enabling automatic label avoidance and more balanced text placement. The final ground-truth image preserves all visible labels, while the input image masks a subset of labels to construct paired samples for historical map completion.

6.5.3 Music Data Synthesis↩︎

We construct image editing and text-to-image tasks for the music domain, covering sheet music, piano roll, and music theory chart representations.

(1) Single-Note Editing and Generation. We programmatically generate single-note data, acting as a simple level, to strengthen perception of fundamental symbols. Editing tasks cover transposition, duration change, accidental addition, and dotting; a notehead-only variant hides the staff to focus on duration symbol transformations. A complementary text-to-image task covers the full combination of pitches, clefs, and durations.

(2) Text-to-Score Generation. This pipeline adopts a three-tier strategy: exhaustive enumeration of fundamental elements (key/time signatures, scales, chords, intervals, etc.), random sampling of valid multi-measure passages with diatonic and stepwise motion constraints, and reverse-engineered note-level descriptions from IrishMAN ABC tunes [55].

(3) Sheet Music Editing. ABC notation folk tunes from IrishMAN are parsed and segmented into measure-level chunks. Deterministic editing operations—transposition, clef change, dynamic marking and ornament insertion—are applied to each chunk and rendered via LilyPond, with pre- and post-edit images forming paired samples.

(4) Piano Roll Editing. MIDI files from Lakh MIDI [56] and MAESTRO [57] are parsed and edited at the note level (transposition, time stretching, rhythmic quantization), then rendered as DAW-style piano roll images. A matrix-level difference check ensures each operation produces visible changes.

(5) Music Theory Chart Editing. Piano keyboard diagrams are programmatically generated, highlighting chord or scale notes on a two-octave keyboard. Editing operations include transposition, chord quality transformation, and mode change across all seven church modes.

(6) Cross-Representation Conversion. Each IrishMAN ABC score segment is simultaneously rendered as both sheet music and piano roll, forming cross-representation conversion pairs.

6.5.4 Computer Science Data Synthesis↩︎

We build a data synthesis engine for the computer science domain using the Python libraries, covering 17 editing and reasoning tasks across three fundamental data structures: trees, graphs, and finite state machines (FSMs). Specifically, the tasks include:

(1) Tree: Topology editing and node manipulation, traversal path visualization, binary search tree (BST) operations, heap operations with dual tree/array views, Huffman coding tree construction, and lowest common ancestor (LCA) finding and path highlighting.

(2) Graph: K-hop neighborhood identification, node degree identification, cycle detection, bipartite graph coloring, shortest path reasoning, and directed graph reachability analysis.

(3) FSM: Input string tracing, state role identification, and transition logic completion.

Rendering engines are selected according to task characteristics, with Matplotlib used for trees and graphs and Graphviz (neato) for FSMs. To preserve spatial consistency between edited image pairs, node layouts are fixed before modification. We further apply automatic quality control to filter invalid or overlapping samples. Chain-of-thought (CoT) annotations are generated synchronously by deterministic solvers and later refined by a large language model (Qwen-Max [84]).

The computer science domain covers 17 visual editing and reasoning tasks organized into three categories: trees, graphs, and finite state machines. All samples are programmatically generated via rule-based engines without reliance on external datasets. Each task supports both the img2img (paired editing images) and text2img (text description to target image) paradigms, producing approximately 64,000 samples per paradigm and roughly 128,000 in total.

Rendering backends are selected based on structural characteristics. Matplotlib is used for tree and general graph tasks because it supports efficient large-scale batch rendering and flexible programmatic control. For finite-state machines (FSMs), we adopt Graphviz with the neato engine to obtain clearer layouts in dense state-transition diagrams.

To maintain spatial consistency within each edited image pair, node coordinates are anchored prior to applying any modification, ensuring that unchanged components preserve identical positions across the before and after images.

We additionally incorporate a geometric collision detection module into the generation pipeline. Node-to-node and node-to-edge distances are explicitly computed, and samples violating predefined minimum-clearance constraints are automatically discarded and regenerated. This procedure substantially improves overall visual quality and reduces overlapping artifacts.

Chain-of-thought (CoT) annotations are produced synchronously using deterministic solvers during data generation and are subsequently rewritten by a large language model (Qwen-Max [84]) to improve readability and linguistic diversity.

(1) Tree Topology Editing. We programmatically generate full binary trees and random general trees, then apply two topology-level operations: completing an incomplete binary tree to a full binary tree of a specified depth, and pruning the subtree rooted at a designated node. Before and after images share the same layout anchors so that unmodified regions remain visually consistent. This task shares a generator with the next one, producing approximately 6,000 pairs combined.

(2) Tree Node Manipulation. Given a randomly generated tree, a parent node satisfying branching-factor constraints is selected and a new child node is inserted with a specified color. The instruction explicitly states the target node ID, parent node, directional placement (left/right child), and node color.

(3) Traversal Path Visualization. Random binary trees are generated and traversed in pre-order, in-order, or post-order. The ground-truth image randomly adopts one of two visualization styles: a smooth spline curve connecting red marker dots along the traversal path, or red circular badges numbering the visit order beneath each node. Approximately 6,000 pairs.

(4) Binary Search Tree Operations. Random BSTs are constructed and subjected to two types of operations: inserting new values with the newly added nodes highlighted by red rectangular boxes, or swapping two node values to violate the BST property, requiring identification and correction of the erroneous positions. The insert-to-fix ratio is approximately 7:3, yielding about 6,000 pairs.

(5) Heap Operations. Random max-heaps or min-heaps are generated, and a single insertion or root extraction is performed. The ground-truth image presents a dual view: a tree view of the heap alongside a bottom array view, with tree nodes and array cells linked by a consistent color mapping. Approximately 4,000 pairs.

(6) Huffman Coding Tree Construction. A set of characters with random frequency weights is generated. The input image displays the scattered leaf nodes, while the ground-truth image shows the complete Huffman tree built by greedy merging, annotated with cumulative weights at internal nodes and 0/1 coding labels on edges. Approximately 2,000 pairs.

(7) Lowest Common Ancestor Finding. Random trees (binary or ternary) are generated, and two descendant nodes from distinct branches are selected. The ground-truth image marks their LCA with a red rectangular box. This task shares a generator with the next one, producing approximately 4,000 pairs combined.

(8) Tree Path Highlighting. On the same type of random tree, two leaf nodes are selected and the ground-truth image highlights the shortest path between them with a thickened red line.

(9) K-Hop Neighborhood Identification. Random connected undirected graphs are generated. Given a start node and distance \(k\), the ground-truth image marks the start node with a gold border and all nodes exactly \(k\) hops away with green borders. Approximately 4,000 pairs.

(10) Node Degree Identification. Given a target degree on a random undirected graph, the ground-truth image highlights all qualifying nodes with green rectangular boxes. Approximately 4,000 pairs.

(11) Cycle Detection. A graph containing exactly one cycle is constructed by adding a single extra edge to a random spanning tree. The ground-truth image highlights all nodes and edges on the cycle in red. Approximately 4,000 pairs.

(12) Bipartite Graph Coloring. Random bipartite graphs are generated with all nodes rendered in grey in the input image. The ground-truth image assigns two distinct colors to the two independent sets such that no adjacent nodes share the same color. Approximately 4,000 pairs.

(13) Shortest Path Reasoning. Given a start and end node on a random connected undirected graph, the ground-truth image highlights all nodes and edges along the BFS shortest path. Approximately 4,000 pairs.

(14) Directed Graph Reachability Analysis. Random directed graphs are generated. Given a start node, the ground-truth image marks all nodes reachable via directed edges with red dashed rectangular borders. Approximately 4,000 pairs.

(15) DFA Input String Tracing. A random deterministic finite automaton (DFA) is generated along with an input string. The ground-truth image highlights the traversed transition edges in blue and fills the final state with green (accepted) or red (rejected). Approximately 4,000 pairs.

(16) State Role Identification. On a randomly generated FSM where all states are initially rendered in grey, the ground-truth image assigns distinct colors to the start state, accept states, and regular states respectively. Approximately 4,000 pairs.

(17) Transition Logic Completion. One transition edge is hidden from a complete DFA. The input image shows the incomplete automaton, and the ground-truth image highlights the restored missing transition in red. Approximately 4,000 pairs.

Tasks 1–14 use NetworkX for topology construction and Matplotlib for rendering; tasks 15–17 use the Graphviz neato engine. All images are rendered with randomized combinations of 14 color themes and multiple font families to increase visual diversity.

6.5.5 Chemistry Data Synthesis↩︎

We construct three types of editing tasks for the chemistry domain.

(1) Molecule Generation. This part of the data is extracted from the ChEMBL database [85]. Text-to-image data are generated from the SMILES [52] formulas of substances using RDKit [50]. We adopt diverse template descriptions together with the given substance name as the prompt, and use the corresponding generated 2D molecular structure image as the ground truth.

(2) Reaction Equation Editing. The data is extracted from USPTO-50K [86] and common chemical reactions from the Internet. Two rendering engines, RDKit [50] and Indigo [51], and seven rendering styles are used to enhance data diversity. The pipeline covers two sub-tasks, both of which use the reactant molecular structures as the input image, while the ground truth is divided into two types: the complete reaction equation and the major product molecule.

(3) Molecular Structure Editing. Molecular structure editing data is constructed based on the ChEMBL database [85] by inserting functional-group fragments into the original molecules to generate editing tasks. The pipeline includes various groups such as phenyl, methoxy, and methylamino, together with insertion validity checking. It supports both single-site editing and multi-site editing with different corresponding positions in the original molecule highlighted in different colors.

6.5.6 Sports Data Synthesis↩︎

We construct the sports portion of the dataset from four subdomains: Chess/Xiangqi, Go, sports nutrition, and soccer tactics. ¹

(1) Chess/Xiangqi: We collect game records from Lichess and XQBase, parse them into structured move sequences, and generate opening, legal_moves, and bestmove editing tasks, where the last task is further filtered by engine-based verification using Stockfish or Pikafish.

Specifically, we construct chess and xiangqi image-editing data from game records. For chess, the raw source is the Lichess standard-game PGN archive. For xiangqi, the raw source is XQBase game pages, which are crawled and parsed into structured JSONL records containing opening names, move sequences, and source identifiers. Based on these structured records, we build three task types: opening, legal_moves, and bestmove. The opening task maps early-game positions to canonical opening states. The legal_moves task highlights all legal destinations for a selected piece. The bestmove task uses external engines, i.e., Stockfish for chess and Pikafish for xiangqi, and keeps only positions whose best moves remain stable under deeper verification and are sufficiently separated from runner-up moves. All samples are rendered into paired editing and gt images, together with a structured JSON file.

(2) Go: We extract candidate board states from KataGo-related SGF archives and further relabel them with KataGo analysis to obtain reliable crucial_move supervision.

Specifically, we construct Go image-editing data from KataGo-related SGF archives. We first parse SGF game trees and sample candidate positions within controlled ply ranges, using the next SGF move as a weak label. We then relabel these candidates with local KataGo analysis to obtain stronger supervision. Positions are retained only when the engine-selected top move is sufficiently reliable, according to thresholds on visits, score gap, and win-rate gap, and when the position is not already too one-sided. The final Go task is crucial_move, where the input is the current board state and the target image adds the key next move marked with “1”.

(3) Sports Nutrition: We build a unified food asset table from USDA FoodData Central together with manually curated GI labels and food images, and render multiple nutrition-oriented editing tasks such as grouping, pyramid construction, pie-chart integration, and attribute highlighting.

Specifically, the sports nutrition pipeline is built from USDA FoodData Central, combined with manually curated food lists, glycemic-index references, and local food images or transparent cutouts. These resources are merged into a unified asset table that stores nutritional attributes, macro categories, GI annotations, and image paths. Based on this asset table, we generate several synthetic image-editing tasks, including classify_grouping, pie_chart_integration, nutrition_pyramid, highlight_high_gi, and highlight_high_protein. The resulting images are fully programmatically rendered while remaining grounded in real nutritional data.

(4) Soccer Tactics: We parse StatsBomb Open Data into canonical tactical records and render formation- and ball-handler-centered editing tasks on synthetic tactics boards. All generated samples are finally converted into a unified multimodal editing format with paired input and target images, textual editing instructions, and structured metadata.

Specifically, the soccer tactics pipeline mainly uses StatsBomb Open Data, with optional auxiliary support from Metrica sample tracking data. We first parse match events and tactical information into a canonical JSONL format containing formations, player coordinates, ball-handler identity, and other metadata. Based on these structured records, we generate several tactics-board editing tasks, including soccer_formation_dots, soccer_formation_jerseys, and soccer_ball_handler_highlight. These tasks do not require external search engines; the supervision is directly derived from structured tactical annotations.

Unified Annotation Format. Although the raw sources and rendering processes differ across sports subdomains, all generated samples are finally converted into a unified multimodal editing format. Each record contains a conditioning image, a target image, a natural-language editing instruction, image size metadata, generation flags, and an original field storing task-specific metadata. Large datasets are automatically sharded into multiple JSONL files, and a corresponding meta.json file records dataset-level information such as the root directory, annotation path, number of shards, and task type.

Instruction Diversification. We further support prompt diversification by rewriting a subset of instructions with semantically equivalent templates while keeping the image supervision unchanged. This increases linguistic variety without altering the underlying visual-editing objective.

7 Training Details↩︎

The reported models use the same training code path in DiffSynth-Studio. Qwen3-VL-8B is used as the reasoning module and is kept frozen. We only optimize LoRA adapters attached to the DiT backbone of the image generator; the original generator weights, text encoder, and VAE remain frozen. All image generators are loaded in bfloat16, and Accelerate is launched with 8 processes on one 8-GPU node. Since the dataloader uses one sample per process and gradient accumulation is 1, the effective batch size is 8 for both T2I generation and editing.

7.0.0.1 Text-to-Image Generation.

The T2I model starts from Qwen-Image-2512 and is trained on the final T2I split of DisciplineGen-1M. We train for one effective pass over this split, corresponding to 43,993 optimizer steps with the effective batch size above. The LoRA rank is 128. We use AdamW with learning rate \(5\times10^{-6}\), weight decay 0, and a constant learning-rate schedule without warmup. Checkpoints are saved every 1,000 optimizer steps and only the latest checkpoint is kept during training. To reduce the tendency to overfit plain white diagram backgrounds, we apply a background loss multiplier of 0.2 to pixels whose RGB distance from the border-estimated background color is at most 24; foreground pixels keep unit weight. The border color is estimated from a 6% border band.

7.0.0.2 Image Editing.

The editing model starts from Qwen-Image-Edit-2511 and is trained on the final editing split of DisciplineGen-1M after validity filtering and duplicate removal. We train for one effective pass, corresponding to 23,384 optimizer steps. The LoRA rank is 64. We use AdamW with learning rate \(2\times10^{-6}\), weight decay 0, and a constant learning-rate schedule without warmup. The editing dataloader uses image,edit_image as the paired image fields and passes edit_image as an extra model input. We enable the Qwen-Image-Edit zero_cond_t option. Checkpointing follows the same setting as T2I: every 1,000 optimizer steps with a single retained checkpoint.

7.0.0.3 Resolution and Data Processing.

For both tasks, we leave the fixed height and width unset and use dynamic-resolution training. Each image is resized/cropped by the dataset operator with a maximum pixel budget of 1,572,864 pixels, and the processed height and width are aligned to multiples of 16. The dataset repeat factor is 1. The training loss is the standard FlowMatch SFT objective used by the Qwen-Image training script; timestep sampling uses 1,000 training timesteps from the pipeline scheduler.

7.0.0.4 LoRA Placement.

For both models, LoRA adapters are inserted into the DiT attention projections, including the query, key, value, added query/key/value, and output projections. For T2I generation, we additionally adapt the image/text MLP output projections and modulation layers. This gives the T2I model more capacity for domain-specific layout, notation, and text-heavy diagram structure, while the editing model uses the more conservative attention-only adapter set to preserve the source image.

References↩︎

---

1. Lichess: https://database.lichess.org/standard/; XQBase: https://www.xqbase.com/xqbase/; KataGo Archive: https://katagoarchive.org/kata1; USDA FoodData Central: https://fdc.nal.usda.gov/; GI labels: https://www.gluok.com/; StatsBomb Open Data: https://github.com/statsbomb/open-data.↩︎