SVGen is an end-to-end model that generates high-quality SVG code from text. We fine-tuned a Large Language Model on our custom SVG-1M dataset using curriculum learning, Chain-of-Thought (CoT), and reinforcement learning.

This repo is built upon LLaMA-Factory. Sincere thanks to their excellent work!

git clone https://github.com/gitcat-404/SVGen.git
cd SVGen

conda create -n svgen python=3.10 -y
conda activate svgen

sudo apt update
sudo apt install libcairo2 libcairo2-dev

pip install torch==2.5.1+cu124 torchvision==0.20.0+cu124 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt
cd LLaMA-Factory && pip install -e ".[torch,metrics]"

For this demonstration, we will be using our top-performing model, SVGen-Qwen2.5-Coder-7B-Instruct. Please download the model weights from Hugging Face and store them under the Models/ path.

pip install huggingface_hub
hf download gitcat-404/SVGen-Qwen2.5-Coder-7B-Instruct --local-dir Models/SVGen-Qwen2.5-Coder-7B-Instruct

python app.py

Taking the test data for this task as an example, we will write the prompts that need to be inferred into a CSV file, such as the example provided in test/color_test.csv.

python inference.py \
    --model_path Models/SVGen-Qwen2.5-Coder-7B-Instruct \
    --csv_file_path test/color_test.csv \
    --prompt_type qwen \
    --output_folder "results/qwen_outputs"

Our evaluation metrics in this article are: Fréchet Inception Distance (FID), CLIPScore-T2I, CLIPScore-I2I, Preference Scores (HPS), and Aesthetic Score. The specific implementation is available in test/ To test the model, first download sac+logos+ava1-l14-linearMSE.pth from Huggingface and hpc.pt from GitHub, placing them in the test/pretrain_weight/ folder. Next, modify test/test.py by filling in the folder containing the previously generated images, then run:"

python test/test.py

All experiments are executed on 8×NVIDIA A800 GPUs using the AdamW optimizer (learning rate 4e-5), with a maximum sequence length of 8,000 tokens.In order to reproduce our experiments, please first download all data from 🤗SVG-1M-Json and place it in the json_data/ folder. We have provided the training scripts in the config/ folder. After modifying the configurations, you can execute them sequentially:

sh config/train_stage1.sh
sh config/train_stage2.sh
sh config/train_stage3.sh
sh config/train_stage_RL.sh

This dataset was collected by web scraping public content from IconFont and is intended for non-commercial academic research and technical exchange purposes only.

python Spider/CrawlingSVG_Icon.py

We would like to extend our sincerest thanks to the projects and websites that inspired this work, specifically: IconFont LLaMA-Factory LLM4SVG OmniSVG Star-vector

@article{wang2025svgen,
  title={SVGen: Interpretable Vector Graphics Generation with Large Language Models},
  author={Wang, Feiyu and Zhao, Zhiyuan and Liu, Yuandong and Zhang, Da and Gao, Junyu and Sun, Hao and Li, Xuelong},
  journal={arXiv preprint arXiv:2508.09168},
  year={2025}
}