LLM4SVG enables Large Language Models to process, understand, and generate complex Scalable Vector Graphics (SVG). It is designed for researchers and developers working with multimodal AI and vector graphics, offering a robust framework for SVG-centric LLM applications.

How It Works

LLM4SVG leverages specialized datasets (SVGX-Core-250k for pretraining, SVGX-SFT-1M for fine-tuning) and integrates with popular LLM frameworks like LLaMA-Factory, Unsloth, Transformers, and TRL. It supports multimodal inputs (text and vision) and offers flexible training options including LoRA and full fine-tuning, with optimized inference via vLLM. The approach includes specialized SVG tokenization for improved SVG syntax handling.

Quick Start & Requirements

• Installation: Clone the repository, create a conda environment (conda env create -f environment.yml && conda activate llm4svg), download datasets (bash script/download_dataset.sh), set up datasets (bash script/setup_dataset.sh), and install LLaMA-Factory (cd LLaMA-Factory && pip install -e ".[torch,metrics]").
• Prerequisites: Python 3.10+, PyTorch, Hugging Face libraries, LLaMA-Factory, Unsloth, TRL, vLLM. GPU acceleration is highly recommended for training and inference.
• Resources: Requires significant GPU memory for training larger models and datasets.
• Links: Project Website, CVPR 2025 Paper, SVGX-Core-250k Dataset, SVGX-SFT-1M Dataset.

Highlighted Details

• Supports fine-tuning a wide range of foundation models including Llama 3.2, Qwen2.5-VL, Gemma 3, DeepSeek, Falcon, Phi-2, and GPT2-XL.
• Offers two specialized datasets: SVGX-Core-250k (pretraining) and SVGX-SFT-1M (supervised fine-tuning).
• Integrates with vLLM for high-throughput, low-latency inference, claiming up to 2x faster generation.
• Provides example configurations for training using LLaMA-Factory, Unsloth, Transformers, and TRL.

Maintenance & Community

The project is associated with the CVPR 2025 conference. For questions, bug reports, or collaboration, users are encouraged to open issues or submit pull requests on GitHub.

Licensing & Compatibility

The project is licensed under the MIT License, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

Pretrained model weights are pending release. The project relies on external frameworks like LLaMA-Factory and Unsloth, which may have their own dependencies and maintenance cycles. Optimal performance may require significant GPU resources and careful configuration of training parameters.