Symbolic Graphics Programming with Large Language Models

^{As training progresses, we can observe that the model acquires better
compositional drawing ability, producing semantically accurate symbolic graphics programs.}

Introduction

LLMs are strong at coding, but their ability to write symbolic graphics programs (SGPs) that render images (especially SVGs) is underexplored. This work studies text-to-SGP generation as a probe of visual generation, introducing SGP-GenBench to evaluate object-, scene-, and composition-level performance across open and proprietary models, revealing notable shortcomings. To improve results, we use reinforcement learning with rewards from visual–text similarity scores, which steadily enhances SVG quality and semantic alignment. Experiments show substantial gains, bringing performance close to state-of-the-art closed-source models.

Installation

conda env create -n sgp_gen -f environment.yml

conda activate sgp_gen

pip install vllm==0.7.2 && pip install oat-llm==0.0.9

git clone git@github.com:Sphere-AI-Lab/SGP-RL.git

cd SGP-RL

pip install -e .

pip install cairosvg openai-clip lxml

We provide a requirements.txt recording the version informations for reference.

Quick Start

Hardware Requirements

We did the experiment with Qwen/Qwen2.5-7B on a 8xH100 node, the mininum requirement is a 8xA100 node.

For the experiments with Qwen/Qwen2.5-3B, we are able to run them on a 8xL40 node.

Prepare Datasets

Automatically Download All Required Data

We provide a script for automatically downloading all training and evaluation datasets:

bash prepare_data.sh

Manually Download datasets

Alternatively, you can also download the datasets manually.

Training Datasets

Setup COCO 2017 dataset(assume you put it in COCO_DIR):

export COCO_DIR=YOUR_COCO_DIR
cd "$COCO_DIR"
# Images
wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip

# Captions annotations
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip train2017.zip
unzip val2017.zip
unzip annotations_trainval2017.zip

You should have:

train2017/
val2017/
annotations/
- captions_train2017.json
- captions_val2017.json ...

setup the svg training data: download the dataset file at https://huggingface.co/datasets/SphereLab/MMSVG-Illustration-40k/resolve/main/MMSVG-Illustration-40k.jsonl

Put it into YOUR_SVG_DIR and setup environment variables:

export SVG_DIR=YOUR_SVG_DIR

SGP-Object Datasets (for evaluation)

Download SGP-Object dataset at
https://huggingface.co/datasets/SphereLab/SGP-Object/resolve/main/SGP-Object.json

Put it into YOUR_SVG_DIR.

RL Training

bash train_zero_svg.sh

Evaluation on SGP-GenBench

Sampling model responses and calculating DINO-score, CLIP-score and Diversity on SGP-:

# modify the sampling parameters to match your training settings
python evaluate_svg_model.py 
    --model_path YOUR_MODEL_PATH

By default, the sampled responses are stored in ./evaluation_results To print out the important metrics as csv,

# modify the sampling parameters to match your training settings
bash print_results.sh

To get the VQA and HPS metrics, check eval_tools

Evaluation on SGP-CompBench: sgp-compbench

Checkpoint

We provide a model checkpoint at training step 780 on HuggingFace and Modelscope:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="SphereLab/SGP-RL")

from modelscope import snapshot_download
model_dir = snapshot_download('NOrangeroli/SGP_RL')

Download via git:

git clone https://huggingface.co/SphereLab/SGP-RL
git clone https://www.modelscope.cn/NOrangeroli/SGP_RL.git

Interactive Inference

We provide a script for interactively run the model:

python interactive_inference.py --model your/model/path

Always prompt in the following form:

Please write SVG code for generating the image corresponding to the following description: YOUR_DESCRIPTION

Acknowledgement

This code is based on understand-r1-zero.

Citation

@article{chen2025symbolic,
  title={Symbolic Graphics Programming with Large Language Models}, 
  author={Yamei Chen and Haoquan Zhang and Yangyi Huang and Zeju Qiu and Kaipeng Zhang and Yandong Wen and Weiyang Liu},
  journal={arXiv preprint arXiv:2509.05208},
  year={2025}
}

bmorphism/SGP-RL