UniMedVL

Paper Abstract

Medical diagnostic applications require models that can process multimodal medical inputs (images, patient histories, lab results) and generate diverse outputs including both textual reports and visual content (annotations, segmentation masks, and images). Despite this need, existing medical AI systems disrupt this unified process: medical image understanding models interpret images but cannot generate visual outputs, while medical image generation models synthesize images but cannot provide textual explanations. This leads to gaps in data representation, feature integration, and task-level multimodal capabilities. To this end, we propose a multi-level framework that draws inspiration from diagnostic workflows through the Observation-Knowledge-Analysis (OKA) paradigm. Specifically, at the observation level, we construct UniMed-5M, a dataset comprising over 5.6M samples that reformat diverse unimodal data into multimodal pairs for foundational observation. At the knowledge level, we propose Progressive Curriculum Learning that systematically introduces medical multimodal knowledge. At the analysis level, we introduce UniMedVL, the first medical unified multimodal model for the simultaneous analysis of image understanding and generation tasks within a single architecture. UniMedVL achieves superior performance on five medical image understanding benchmarks, while matching specialized models in generation quality across eight medical imaging modalities. Crucially, our unified architecture enables bidirectional knowledge sharing: generation tasks enhance visual understanding features, demonstrating that integrating traditionally separate capabilities within a single medical framework unlocks improvements across diverse medical vision-language tasks. Code is available at this https URL .

📚 Introduction

We introduce UniMedVL, the unified medical foundation model for seamless multimodal understanding and generation. Four key innovations distinguish UniMedVL:

• Unified Observation-Knowledge-Analysis Architecture: UniMedVL sets itself apart from prior medical AI models by following a clinically-inspired three-level framework that mirrors how physicians process medical information, enabling both understanding and generation within a single architecture.

• Versatile Medical Multimodal Capabilities: UniMedVL supports a broad spectrum of medical tasks, including visual question answering, medical report generation, text-to-medical-image synthesis, cross-modal translation, and virtual staining across 9 imaging modalities.

• Large-Scale Medical Dataset: We present UniMed-5M, a comprehensive medical multimodal dataset containing 5.6M+ high-quality samples with three-stage quality verification and expert validation, covering understanding, generation, and interleaved tasks.

• Superior Performance: UniMedVL achieves state-of-the-art performance on multiple evaluation datasets, with 75.4% accuracy on SLAKE VQA, 53.5% on PathVQA, and competitive generation quality (96.29 average gFID), setting a new standard in unified medical AI.

🔬 Methodology

📋 OKA Framework: Observation-Knowledge-Analysis

UniMedVL follows a workflow-guided three-level framework that mirrors how physicians process medical information:

flowchart TD
    A[Observation Level] --> B[Knowledge Level] --> C[Analysis Level]

    A1[UniMed-5M Dataset<br/>5.6M samples<br/>8 imaging modalities] --> A
    A --> A2[Quality Control<br/>Three-stage verification<br/>Expert validation]

    B1[Progressive Curriculum<br/>Foundation → Instruction → Unified] --> B
    B --> B2[Cross-modal Knowledge Fusion<br/>Understanding ↔ Generation]

    C1[Unified Architecture<br/>Dual encoders + MOT] --> C
    C --> C2[Multimodal Outputs<br/>Reports + Images + Annotations]

🎯 Training Strategy

Three-Stage Progressive Curriculum Learning:

1. 🔧 Stage 1 - Foundation Training (85K steps)

   • Basic medical pattern recognition
   • Visual-language alignment
   • Data ratio: 75% I2T, 25% T2I

2. 📚 Stage 2 - Instruction Tuning (120K steps)

   • Cross-modal understanding enhancement
   • Medical expertise development
   • Data ratio: 40% I2T, 45% T2I, 10% Interleaved

3. 🚀 Stage 3 - Unified Training (70K steps)

   • Advanced multimodal synthesis
   • Interleaved task mastery
   • Data ratio: 37% I2T, 35% T2I, 25% Interleaved

Model Details

Model Description

UniMedVL unifies medical multimodal understanding and generation within a single 14B-parameter architecture. The model supports visual question answering, medical report generation, text-to-medical-image synthesis, cross-modal translation, and virtual staining across 9 imaging modalities (CXR, CT, MRI, Ultrasound, Histopathology, Retinal Fundus, OCT, Endoscopy).

• License: Apache License 2.0
• Model Size: 14B parameters

Model Sources

• Repository: https://github.com/uni-medical/UniMedVL
• Project Page: https://uni-medical.github.io/UniMedVL_Web/
• Paper: https://arxiv.org/abs/2510.15710

Uses

Direct Use

The model can be directly used for:

• Medical Visual Question Answering: Answer clinical questions about medical images
• Medical Report Generation: Generate radiology reports from medical images
• Text-to-Medical-Image Synthesis: Generate medical images from textual descriptions
• Cross-Modal Translation: Convert between different medical imaging modalities
• Virtual Staining: Transform H&E images to IHC staining

Out-of-Scope Use, Cautious!!

• Clinical Decision Making: This model is for research purposes only and should NOT be used for actual clinical diagnosis or treatment decisions

💬 Qualitative Results

Here we present some comprehensive visualization results demonstrating UniMedVL's capabilities. For additional visualization results and comparisons, please see our Project Page.

Chest X-Ray (CXR)

Computed Tomography (CT)

Magnetic Resonance Imaging (MRI)

Ultrasound

Histopathology (HIS)

Retinal Fundus Photography (CFP)

Optical Coherence Tomography (OCT)

Endoscopy

📊 Quantitative Performance

Text-to-image generation performance across 8 medical imaging modalities. Metrics: gFID ↓ (lower is better) / BioMedCLIP Score ↑ (higher is better)

Virtual Immunohistochemistry Staining (H&E → IHC)

MRI Super-Resolution (4× upsampling)

Cross-Modal Synthesis (T2 ↔ FLAIR MRI)

Performance on counterfactual chest X-ray generation with explanatory text. † indicates unified fine-tuning variant.

📝 Open-Source Plan

• 📄 Paper & Evaluations - Research documentation and evaluation results
• 🖼️ Visualizations - Result figures and model demonstrations
• 💾 Model Checkpoints - Pre-trained UniMedVL weights (14B parameters)
• 🔧 Inference Code - Model loading and inference examples
• 🏋️ Training Code - Full training pipeline and configuration files
• 📁 UniMed-5M Dataset - Training dataset with quality control

🚀 Getting Started

Installation

conda env create -f codes/environment.yaml
conda activate unimedvl

Inference Scripts

Two interactive inference scripts are provided in the codes/ directory:

1. Medical Visual Question Answering (interactive_vqa_inferencer.py)

2. Medical Image Generation (interactive_image_generator.py)

Quick Usage

1. Download the UniMedVL checkpoint
2. Set model_path and ROOT in the script configuration
3. Run the script: python codes/interactive_vqa_inferencer.py or python codes/interactive_image_generator.py

📜 License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

📚 Citations

If you use this project in your research or work, please cite it as:

 @misc{ning2025unimedvlunifyingmedicalmultimodal,
      title={Unimedvl: Unifying Medical Multimodal Understanding And Generation Through Observation-Knowledge-Analysis}, 
      author={Junzhi Ning and Wei Li and Cheng Tang and Jiashi Lin and Chenglong Ma and Chaoyang Zhang and Jiyao Liu and Ying Chen and Shujian Gao and Lihao Liu and Yuandong Pu and Huihui Xu and Chenhui Gou and Ziyan Huang and Yi Xin and Qi Qin and Zhongying Deng and Diping Song and Bin Fu and Guang Yang and Yuanfeng Ji and Tianbin Li and Yanzhou Su and Jin Ye and Shixiang Tang and Ming Hu and Junjun He},
      year={2025},
      eprint={2510.15710},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.15710}, 

🙏 Acknowledgments

We sincerely thank the following projects and their contributors for their invaluable open-source contributions that made this research possible:

• Bagel - Foundation model architecture and training methodology inspiration
• HealthGPT - Medical domain adaptation and evaluation framework
• VLMEvalKit - Comprehensive evaluation toolkit for vision-language models