DisciplineGen-1M: A Large-Scale Dataset for Multidisciplinary Visual Generation and Editing

2026-07-02 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

DisciplineGen-1M is a new million-scale multidisciplinary dataset designed to improve image generation and editing for knowledge-intensive diagrams. It comprises 1.2 million samples across ten disciplines, including mathematics, physics, chemistry, biology, geography, computer science, economics, history, music, and sports. The dataset was constructed using a scalable framework that integrates vector-graphics rendering, OCR-based editing, curated programmatic synthesis, and large-scale text-to-image filtering. This process generates captions, editing instructions, structured annotations, and paired images with controllable semantic differences. Building on DisciplineGen-1M, a discipline-informed reasoning-generation model was developed, demonstrating substantial improvements over open-source baselines on discipline-related benchmarks like GenExam and GRADE. Evaluations on general reasoning benchmarks, WISE and RISE, also indicated broader transfer capabilities. The findings highlight that large-scale structured academic visual data is essential for advancing image generation from aesthetic plausibility to verifiable, knowledge-grounded visual creation. The dataset, model, and data curation pipeline source code will be publicly released.

Key takeaway

For computer vision engineers developing models for knowledge-intensive visual generation, you should integrate DisciplineGen-1M into your training pipelines. This dataset's 1.2 million multidisciplinary samples and structured annotations offer a direct path to improving model reliability and verifiable knowledge-grounded output. Consider utilizing the publicly released dataset and model to move beyond aesthetic plausibility towards precise, discipline-informed visual creation in your applications.

Key insights

Large-scale structured academic visual data is key for knowledge-grounded image generation and editing.

Principles

• Knowledge-intensive diagrams require disciplinary concepts.
• Precise spatial relations are critical for correctness.
• Structured academic data improves reasoning-generation.

Method

DisciplineGen-1M's construction combines vector-graphics rendering, OCR-based editing, curated programmatic synthesis, and large-scale text-to-image filtering to produce diverse visual data with annotations.

In practice

• Use DisciplineGen-1M for diagram generation.
• Apply OCR-based editing for visual data.
• Integrate structured annotations for model training.

Topics

• DisciplineGen-1M
• Image Generation
• Image Editing
• Multidisciplinary Datasets
• Knowledge-Grounded AI
• Computer Vision

Best for: AI Scientist, Computer Vision Engineer, Research Scientist

Related on AIssential

• DeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing — DeepGen 1.0 is a lightweight 5B multimodal model outperforming larger counterparts in image generation and editing.
• Knowledge-Intensive Video Generation — Knowledge-intensive video generation requires new benchmarks and metrics to evaluate factuality and practical usefulness.
• ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding — Code-guided synthetic data generation significantly improves VLM chart understanding and reasoning capabilities.
• GPIC: A Giant Permissive Image Corpus for Visual Generation — GPIC provides a massive, permissively licensed, and safety-filtered image corpus for scalable visual generative modeling research.
• Image Generators are Generalist Vision Learners — Instruction-tuning image generators unlocks powerful generalist visual understanding capabilities, mirroring LLM emergent behaviors.
• Paper Digest: ICCV 2025 Papers & Highlights — Multimodal AI, 3D reconstruction, and video understanding are advancing rapidly with novel models and benchmarks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.