Technical Summary: DisciplineGen-1M

Problem Statement

Current image generation and editing models have achieved remarkable success in producing aesthetically pleasing natural images. However, they remain unreliable when tasked with generating or editing knowledge-intensive diagrams (e.g., in mathematics, physics, chemistry, biology, etc.). In these disciplinary contexts, correctness depends on precise symbolic structures, spatial relations, and academic concepts rather than just visual plausibility. A minor error in a symbol, label, or molecular group can invalidate the entire result.

Existing datasets for image generation and editing largely focus on natural images, artworks, or general world knowledge (e.g., WISE, RISE). While some specialized datasets exist (e.g., StructVisuals, TextAtlas5M), they often lack multidisciplinary scope, joint text-to-image (T2I) and editing supervision, or explicit reasoning and knowledge annotations. There is a critical shortage of large-scale, structured datasets designed to train models to handle the rigorous constraints of academic visual creation.

Methodology

1. Data Construction: DisciplineGen-1M

The authors introduce DisciplineGen-1M, a million-scale dataset containing 1.2 million samples spanning ten disciplines: mathematics, physics, chemistry, biology, geography, computer science, economics, history, music, and sports. The dataset supports both T2I generation and image editing.

To address the heterogeneity of academic images, the authors designed a scalable framework combining four complementary pipelines:

• Vector Graphics Rendering (SVG & TikZ):
  • SVG Branch: Used for biology, physics, and economics. It constructs structured semantic instances (nodes, relations, curves) mapped to templates and rendered as SVG. This allows for deterministic re-rendering and explicit control over semantic edits (e.g., removing labels, shifting curves). It also utilizes web image initialization where Gemini-3-Flash generates SVG code from candidate images.
  • TikZ Branch: Targets biology, geography, physics, chemistry, CS, and economics. It uses a seed library of 1,346 knowledge points expanded into drawing blueprints and translated into compilable TikZ code. A "compiler-in-the-loop" repair process fixes LaTeX errors, and a Vision-Language Model (VLM) judge filters for correctness, clarity, and completeness. Editing is achieved by replacing labels with placeholders and recompiling.
• OCR-Based Editing:
  • Targets educational figures with informative labels. Using PP-OCR v5, text instances and bounding boxes are extracted. A VLM selects regions suitable for masking (rejecting ambiguous or unimportant text). The masked image serves as input, the original as ground truth, and the VLM generates the editing instruction.
• Large-Scale T2I Data Filtering:
  • Filters open-source educational image pools (e.g., K12, MV-MATH). The pipeline applies pHash deduplication, resolution filtering, and OCR pre-screening to remove text-dominant or language-dependent images. Remaining images are scored by Qwen3-VL-8B on dimensions like discipline relevance, diagrammatic nature, visual complexity, and domain knowledge.
• Specialized Programmatic Synthesis:
  • Uses domain-specific engines to synthesize data where web data is sparse.
    • Math: Matplotlib, MathCanvas, GeoGebra for function plots, geometry, and solid transformations.
    • Chemistry: RDKit and Indigo for molecule generation and reaction equations.
    • Computer Science: Graphviz and Matplotlib for trees, graphs, and finite-state machines.
    • History: Wikimedia Commons and basemaps for timelines and maps.
    • Music: IrishMAN ABC tunes and LilyPond for sheet music and piano rolls.
    • Sports: Chess/Xiangqi records, Go SGF, and nutrition data for board games and tactics.

2. Model Architecture

Building on DisciplineGen-1M, the authors introduce a discipline-informed reasoning-generation model:

• Reasoning Module: Qwen3-VL-8B converts implicit disciplinary requests into explicit reasoning plans, identifying relevant knowledge, target elements, and structural changes.
• Generation Modules:
  • T2I: Qwen-Image-2512.
  • Editing: Qwen-Image-Edit-2511.
• Training Strategy: The models are adapted using LoRA (Low-Rank Adaptation) on the DiT backbone. The training utilizes a FlowMatch SFT loss with dynamic-resolution loading. The reasoning module is kept frozen, while the image generators are fine-tuned.

Key Contributions

1. Scalable Framework: A novel data construction framework combining vector rendering, OCR-based editing, programmatic synthesis, and large-scale filtering to produce high-quality, structured academic visuals.
2. DisciplineGen-1M Dataset: The first million-scale (1.2M) multidisciplinary visual generation dataset, covering T2I and editing across ten distinct academic fields with structured supervision.
3. State-of-the-Art Performance: Demonstration that adapting open-source generators with this dataset yields SOTA results among open-source models on discipline-specific benchmarks, with significant generalization to broader reasoning tasks.

Experimental Results

Discipline-Specific Benchmarks

• GenExam (T2I Generation): The model achieved an overall relaxed score of 51.4, ranking first among open-source models. It outperformed Qwen-Image-2512 by 15.6 points and FLUX.2 dev by 8.5 points. Improvements were consistent across all ten disciplines, with notable gains in chemistry, economics, music, CS, and history.
• GRADE (Image Editing): The editing model reached an overall relaxed score of 58.7, surpassing the best open-source baseline (DreamOmni) by 14.4 points and Qwen-Edit-2511 by 28.7 points. Gains were broad across physics, chemistry, math, geography, biology, CS, and history.

General Reasoning Benchmarks

• WISE (World-Knowledge T2I): The model scored 0.66, outperforming Qwen-Image-2512 by 0.16. Gains were strongest in spatial, biology, physics, and chemistry categories, indicating transfer of structural knowledge to world-knowledge tasks involving scientific constraints.
• RISE (Reasoning-Informed Editing): The model achieved 26.4 accuracy, outperforming Qwen-Image-Edit-2511 by 7.0 points and BAGEL (thinking) by 10.8 points. It notably exceeded the strongest closed-source baseline in overall score, driven by improvements in temporal, causal, and spatial reasoning.

Qualitative Analysis

Visual comparisons show that the proposed model better preserves key structural constraints (e.g., molecular connectivity, graph topology, musical notation) compared to baselines, which often produce visually plausible but structurally incorrect outputs.

Significance and Claims

The paper claims that large-scale structured academic visual data is a key ingredient for moving image generation from "aesthetic plausibility" toward "verifiable knowledge-grounded visual creation."

The authors assert that:

1. Training on multidisciplinary, structured data improves not only visual fidelity but also the ability to follow discipline-specific constraints (e.g., symbolic notation, spatial relations).
2. The benefits of this structured supervision extend beyond specific academic domains to general reasoning-informed generation and editing, particularly in tasks requiring spatial, causal, or scientific reasoning.
3. While a gap remains between open-source and closed-source models, DisciplineGen-1M significantly narrows this gap, validating the utility of explicit knowledge constraints and paired structured supervision for academic visual creation.

The authors publicly release the dataset, the trained models, and the source code for the data curation pipeline to ensure reproducibility.