InternSVG - A unified SVG modeling suite launched by Shanghai AI Lab

What is InternSVG?

InternSVG is a comprehensive "data-evaluation-model" suite for unified SVG modeling, developed by the Shanghai Artificial Intelligence Laboratory and other institutions. It comprises three main components: the SAgoge dataset with over 16 million samples, the standardized SArena benchmark, and a unified multimodal large-scale model based on InternVL3-8B. Through SVG-specific tokens and two-stage training, InternSVG achieves unified modeling for the understanding, editing, and generation of icons, illustrations, chemical structures, and animations, significantly surpassing existing methods.

Main functions of InternSVG

• Semantic understandingInternSVG can parse the semantics and structure of SVG code, support the generation of detailed descriptions and answers to multiple-choice questions, and accurately identify graphic content and attributes.
• Command EditingIt supports 10 editing operations, covering low-level editing such as color modification and geometric transformation, as well as advanced editing such as semantic color replacement and style transfer.
• Cross-modal generationSupports generating static SVGs (icons, illustrations, chemical formulas) from text or images, as well as generating vector animations from text or video.

The technical principles of InternSVG

• Architectural foundationThe ViT-MLP-LLM paradigm is built based on InternVL3-8B. InternViT-300M is used as a visual encoder to process input images or videos, and the Qwen2.5-7B language model is connected through the MLP projection layer for sequence modeling.
• SVG Exclusive TokenMore than 200 special tokens were designed for SVG syntax, covering 55 core tags, 42 attributes and numerical ranges, compressing the sequence length by 30-50% and effectively alleviating the context pressure of long sequence modeling.
• Subword embedding initializationThe new token is decomposed into pre-trained subwords and their embeddings are averaged as initial values, preserving semantic prior knowledge and reducing the training loss by about 40%, which significantly accelerates convergence.
• Two-stage progressive trainingThe first stage trains the model on simple icons and chemical data to establish basic grammatical cognition. The second stage introduces long sequence illustrations and complex animation data, and gradually improves the model's ability to handle complex structures through a course learning strategy.
• Unified Task ModelingBy using a shared Transformer architecture to handle three types of tasks simultaneously—understanding, editing, and generating—it enables positive transfer of knowledge across tasks, avoids training models separately for each task, and significantly improves parameter efficiency and generalization ability.

How to use InternSVG

• Environmental preparationAfter cloning the repository, create a Python 3.9 virtual environment and install the dependencies. If you need to evaluate, download the ViCLIP checkpoint.
• Model DeploymentDownload the InternSVG-8B model weights from HuggingFace and launch the API service with LMDeploy to support multi-GPU parallel inference.
• Invoking inference: Send requests via the standard OpenAI API format, supporting input of text or images to generate SVG, or input of SVG code for understanding and editing.
• Custom TrainingPrepare the dataset and run the script that adds a special token, then perform fine-tuning training in the first stage (simple data) and the second stage (full data) in sequence.
• Model EvaluationDownload the SArena benchmark dataset, perform inference on the model output, and calculate various evaluation metrics to verify performance.

InternSVG project address

• GitHub repositoryhttps://github.com/hmwang2002/InternSVG
• HuggingFace model libraryhttps://huggingface.co/InternSVG/InternSVG-8B
• arXiv technical paperhttps://arxiv.org/pdf/2510.11341

Key information and usage requirements for InternSVG

• Project PositioningInternSVG is a unified SVG modeling suite developed by the Shanghai Artificial Intelligence Laboratory in collaboration with Shanghai Jiao Tong University, Nanjing University, and other institutions. It has been accepted by ICLR 2026.
• Core componentsIt comprises three core components: the SAgoge multimodal dataset with over 16 million samples (covering four major areas: icons, illustrations, chemical structures, and vector animations), the standardized SArena comprehensive evaluation benchmark (providing unified task definitions and evaluation metrics), and the InternSVG-8B unified multimodal large model based on InternVL3-8B.
• Hardware environmentInference requires at least a single GPU (multiple GPUs are recommended to improve throughput), while training requires 96 NVIDIA A800 GPUs or equivalent computing power.
  Software DependencyPython 3.9 is required, and PyTorch, Transformers, LMDeploy (for service deployment) and LLaMA-Factory (for training) need to be installed.

InternSVG's core advantages

• Unified modeling across all tasksBreaking away from the limitations of traditional SVG understanding, editing, and generation tasks being isolated from each other, it achieves cross-task knowledge positive transfer through a single model architecture, avoiding repeated model training for different tasks and significantly improving parameter efficiency and generalization ability.
• Ultra-large-scale data foundationBased on SAgoge, the largest SVG multimodal dataset to date (with over 16 million samples), covering four high-value areas—icons, long sequence illustrations, chemical structures, and vector animations—it provides comprehensive training support for models, ranging from simple static graphics to complex dynamic sequences.
• Professional and technical optimizationThe system innovatively introduces over 200 SVG-specific special tokens and a sub-word embedding initialization strategy, coupled with a two-stage progressive training (from simple icons to complex animations), effectively compressing sequence length by 30-50%, accelerating convergence, and significantly reducing the difficulty of modeling long sequences.
• Leading in overall performanceIn the SArena standardized benchmark test, compared with the strongest proprietary models such as Claude-4-Sonnet and GPT-4o, the understanding accuracy is improved by about 11%, the PSNR of the editing task is improved by about 34%, and the FID of the generation task is reduced by about 56%, achieving significant superiority across all tasks and domains.
• Precision in professional fieldsIn professional scenarios such as the generation of chemical organic structural formulas, where topological accuracy and symbol standardization are extremely important, the accuracy of bond angle, atom labeling, and ring structure reduction is significantly higher than existing baselines, meeting the high-precision requirements of scientific visualization.

Comparison of InternSVG with similar competitors

Application scenarios of InternSVG

• Digital Design CreationDesigners can quickly generate high-quality vector icons and illustrations using text descriptions or reference images. It supports fine-grained editing such as color adjustment and style transfer based on natural language commands, significantly improving UI/UX design efficiency.
• Scientific VisualizationIn scientific research fields such as chemistry and biology, researchers can automatically convert molecular names or structural images into precise vector chemical structural formulas, meeting the strict requirements for graphic standardization and editability in paper illustrations and textbook publications.
• Intelligent content understandingAutomatically parses the semantic content and geometric structure of SVG icons or illustrations, supporting visual impairment assistance, image retrieval, and intelligent question answering and knowledge extraction based on graphic content.
• Dynamic Media ProductionGenerate vector animations (such as loading animations and icon effects) from text scripts or reference videos, outputting lightweight and infinitely scalable animation resources to adapt to multi-terminal display needs.
• Education and Technology DocumentsInternSVG supports real-time adjustment of graphic elements via natural language commands, lowering the barrier for non-design professionals to create high-quality vector teaching content.