star-vector

🔥 News

• Sep 2025: RLRF Accepted at NeurIPS 2025,
  • Our follow-up work to StarVector, RLRF, has been accepted to NeurIPS 2025! Check out the paper [Link]
• March 2025: StarVector Accepted at CVPR 2025,
  • StarVector has been accepted at CVPR 2025! Check out the paper [Link]
  • Check out our website for more information [Link]
  • StarVector models are now available on HuggingFace! [Link] [Link]
  • SVGBench and SVG-Stack datasets are now available on HuggingFace Datasets! [Link] [Link]

🚀 Introduction

StarVector is a multimodal vision-language model for Scalable Vector Graphics (SVG) generation. It can be used to perform image2SVG and text2SVG generation. We pose image generation as a code generation task, using the power of multimodal VLMs

Abstract: Scalable Vector Graphics (SVGs) are vital for modern image rendering due to their scalability and versatility. Previous SVG generation methods have focused on curve-based vectorization, lacking semantic understanding, often producing artifacts, and struggling with SVG primitives beyond \textit{path} curves. To address these issues, we introduce StarVector, a multimodal large language model for SVG generation. It performs image vectorization by understanding image semantics and using SVG primitives for compact, precise outputs. Unlike traditional methods, StarVector works directly in the SVG code space, leveraging visual understanding to apply accurate SVG primitives. To train StarVector, we create SVG-Stack, a diverse dataset of 2M samples that enables generalization across vectorization tasks and precise use of primitives like ellipses, polygons, and text. We address challenges in SVG evaluation, showing that pixel-based metrics like MSE fail to capture the unique qualities of vector graphics. We introduce SVG-Bench, a benchmark across 10 datasets, and 3 tasks: Image-to-SVG, Text-to-SVG generation, and diagram generation. Using this setup, StarVector achieves state-of-the-art performance, producing more compact and semantically rich SVGs.

Multimodal Architecture

StarVector uses a multimodal architecture to process images and text. When performing Image-to-SVG (or image vectorization), the image is projected into visual tokens, and SVG code is generated. When performing Text-to-SVG, the model only receives the text instruction (no image is provided), and a novel SVG is created. The LLM is based of StarCoder, which we leverage to transfer coding skills to SVG generation.

📖 Table of Contents

• 💿 Installation
• 🏎️ Quick Start - Image2SVG Generation
• 🎨 Models
• 📊 Datasets
• 🏋️‍♂️ Training
• 🏆 Evaluation on SVG-Bench
• 🧩 Demo
• 📚 Citation
• 📝 License

Installation

1. Clone this repository and navigate to star-vector folder

git clone https://github.com/joanrod/star-vector.git
cd star-vector

2. Install Package

conda create -n starvector python=3.11.3 -y
conda activate starvector
pip install --upgrade pip  # enable PEP 660 support
pip install -e .

3. Install additional packages for training

pip install -e ".[train]"

Upgrade to latest code base

git pull
pip install -e .

Quick Start - Image2SVG Generation

from PIL import Image
from starvector.model.starvector_arch import StarVectorForCausalLM
from starvector.data.util import process_and_rasterize_svg

model_name = "starvector/starvector-8b-im2svg"

starvector = StarVectorForCausalLM.from_pretrained(model_name)

starvector.cuda()
starvector.eval()

image_pil = Image.open('assets/examples/sample-0.png')
image = starvector.process_images([image_pil])[0].cuda()
batch = {"image": image}

raw_svg = starvector.generate_im2svg(batch, max_length=1000)[0]
svg, raster_image = process_and_rasterize_svg(raw_svg)

Use it from HuggingFace AutoModel

from PIL import Image
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor
from starvector.data.util import process_and_rasterize_svg
import torch

model_name = "starvector/starvector-8b-im2svg"

starvector = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, trust_remote_code=True)
processor = starvector.model.processor
tokenizer = starvector.model.svg_transformer.tokenizer

starvector.cuda()
starvector.eval()

image_pil = Image.open('assets/examples/sample-18.png')

image = processor(image_pil, return_tensors="pt")['pixel_values'].cuda()
if not image.shape[0] == 1:
    image = image.squeeze(0)
batch = {"image": image}

raw_svg = starvector.generate_im2svg(batch, max_length=4000)[0]
svg, raster_image = process_and_rasterize_svg(raw_svg)

Models

We provide Hugging Face 🤗 model checkpoints for image2SVG vectorization, for 💫 StarVector-8B and 💫 StarVector-1B. These are the results on SVG-Bench, using the DinoScore metric.

Method	SVG-Stack	SVG-Fonts	SVG-Icons	SVG-Emoji	SVG-Diagrams
AutoTrace	0.942	0.954	0.946	0.975	0.874
Potrace	0.898	0.967	0.972	0.882	0.875
VTracer	0.954	0.964	0.940	0.981	0.882
Im2Vec	0.692	0.733	0.754	0.732	-
LIVE	0.934	0.956	0.959	0.969	0.870
DiffVG	0.810	0.821	0.952	0.814	0.822
GPT-4-V	0.852	0.842	0.848	0.850	-
💫 StarVector-1B (🤗 Link)	0.926	0.978	0.975	0.929	0.943
💫 StarVector-8B (🤗 Link)	0.966	0.982	0.984	0.981	0.959

Note: StarVector models will not work for natural images or illustrations, as they have not been trained on those images. They excel in vectorizing icons, logotypes, technical diagrams, graphs, and charts.

Datasets - SVG-Bench

SVG-Bench is a benchmark for evaluating SVG generation models. It contains 10 datasets, and 3 tasks: Image-to-SVG, Text-to-SVG, and Diagram-to-SVG.

See our Huggingface 🤗 Dataset Collection

Dataset	Train	Val	Test	Token Length	SVG Primitives	Annotation
SVG-Stack (🤗 Link)	2.1M	108k	5.7k	1,822 ± 1,808	All	Captions
SVG-Stack_sim (🤗 Link)	601k	30.1k	1.5k	2k ± 918	Vector path	-
SVG-Diagrams (🤗 Link)	-	-	472	3,486 ± 1,918	All	-
SVG-Fonts (🤗 Link)	1.8M	91.5k	4.8k	2,121 ± 1,868	Vector path	Font letter
SVG-Fonts_sim (🤗 Link)	1.4M	71.7k	3.7k	1,722 ± 723	Vector path	Font letter
SVG-Emoji (🤗 Link)	8.7k	667	668	2,551 ± 1,805	All	-
SVG-Emoji_sim (🤗 Link)	580	57	96	2,448 ± 1,026	Vector Path	-
SVG-Icons (🤗 Link)	80.4k	6.2k	2.4k	2,449 ± 1,543	Vector path	-
SVG-Icons_sim (🤗 Link)	80,435	2,836	1,277	2,005 ± 824	Vector path	-
SVG-FIGR (🤗 Link)	270k	27k	3k	5,342 ± 2,345	Vector path	Class, Caption

We offer a summary of statistics about the datasets used in our training and evaluation experiments. This datasets are included in SVG-Bench. The subscript sim stands for the simplified version of the dataset, as required by some baselines.

Training

Confirm dependencies are installed

pip install -e ".[train]"

Set environment variables

We recommend setting the following environment variables:

  export HF_HOME=<path to the folder where you want to store the models>
  export HF_TOKEN=<your huggingface token>
  export WANDB_API_KEY=<your wandb token>
  export OUTPUT_DIR=<path/to/output>

cd the root of the repository.

cd star-vector

Image2SVG Pretraining (Stage 1)

We have different training approaches for StarVector-1B and StarVector-8B. StarVector-1B can be trained using Deepspeed, while StarVector-8B requires FSDP.

StarVector-1B Training

You can use the following command to train StarVector-1B on SVG-Stack for the Image2SVG vectorization task, using Deepspeed and Accelerate

# StarVector-1B
accelerate launch --config_file configs/accelerate/deepspeed-8-gpu.yaml starvector/train/train.py config=configs/models/starvector-1b/im2svg-stack.yaml

StarVector-8B Training

You can use the following command to train StarVector-8B on SVG-Stack for the Image2SVG vectorization task, using FSDP and Accelerate. We provide the torchrun command to support multi-nodes and multi-GPUs.

# StarVector-8B
torchrun \
  --nproc-per-node=8 \
  --nnodes=1 \
  starvector/train/train.py \
  config=configs/models/starvector-8b/im2svg-stack.yaml

Finetuning StarVector (Stage 2)

After pretraining StarVector on image vectorization, we finetune it on additional SVG tasks like Text2SVG, and SVG-Bench datasets.

Text2SVG Finetuning

# StarVector-1B
accelerate launch --config_file config/accelerate/deepspeed-8-gpu.yaml starvector/train/train.py config=configs/models/starvector-1b/text2svg-stack.yaml

# StarVector-8B
torchrun \
  --nproc-per-node=8 \
  --nnodes=1 \
  starvector/train/train.py \
  config=configs/models/starvector-8b/text2svg-stack.yaml

SVG-Bench Finetuning

# StarVector-1B
accelerate launch --config_file config/accelerate/deepspeed-8-gpu.yaml starvector/train/train.py config=configs/models/starvector-1b/im2svg-{fonts,icons,emoji}.yaml

# StarVector-8B
torchrun \
  --nproc-per-node=8 \
  --nnodes=1 \
  starvector/train/train.py \
  config=configs/models/starvector-8b/im2svg-{fonts,icons,emoji}.yaml

We also provide shell scripts in scripts/train/*

Validation on SVG Benchmarks (⭐ SVG-Bench)

We validate StarVector on ⭐ SVG-Bench Benchmark. We provide the SVGValidator class that allows you to run StarVector using 1) the HuggingFace generation backend or 2) the VLLM backend. The later is substantially faster thanks to the use of Paged Attention.

HuggingFace Generation Backend

Let's start with the evaluation for StarVector-1B and StarVector-8B on SVG-Stack, using the HuggingFace generation backend (StarVectorHFAPIValidator). To override the input arguments, you can add cli args following the yaml file structure.

# StarVector-1B on SVG-Stack, using the HuggingFace backend 
python starvector/validation/validate.py \
config=configs/generation/hf/starvector-1b/im2svg.yaml \
dataset.dataset_name=starvector/svg-stack

# StarVector-8B on SVG-Stack, using the vanilla HuggingFace generation API
python starvector/validation/validate.py \
config=configs/generation/hf/starvector-8b/im2svg.yaml \
dataset.dataset_name=starvector/svg-stack

vLLM Backend

For using the vLLM backend (StarVectorVLLMAPIValidator), first install our StarVector fork of VLLM, here.

git clone https://github.com/starvector/vllm.git
cd vllm
pip install -e .

Then, launch the using the vllm config file (it uses StarVectorVLLMValidator):

# StarVector-1B
python starvector/validation/validate.py \
config=configs/generation/vllm/starvector-1b/im2svg.yaml \
dataset.dataset_name=starvector/svg-stack

# StarVector-8B
python starvector/validation/validate.py \
config=configs/generation/vllm/starvector-8b/im2svg.yaml \
dataset.dataset_name=starvector/svg-stack

We provide evaluation scripts in scripts/eval/*

StarVector Demo

The demo provides two options for converting images to SVG code:

1. HuggingFace generation functionality
2. VLLM (recommended) - offers faster generation speed

Option 1: HuggingFace Generation with Gradio Web UI

We provide a Gradio web UI for you to play with our model.

Launch a controller

python -m starvector.serve.controller --host 0.0.0.0 --port 10000

Launch a gradio web server.

python -m starvector.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload --port 7000

You just launched the Gradio web interface. Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker.

Launch a model worker

This is the actual worker that performs the inference on the GPU. Each worker is responsible for a single model specified in --model-path.

python -m starvector.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path joanrodai/starvector-1.4b

Wait until the process finishes loading the model and you see "Uvicorn running on ...". Now, refresh your Gradio web UI, and you will see the model you just launched in the model list.

You can launch as many workers as you want, and compare between different model checkpoints in the same Gradio interface. Please keep the --controller the same, and modify the --port and --worker to a different port number for each worker.

vllm serve starvector/starvector-8b-im2svg --chat-template configs/chat-template.jinja --trust-remote-code --port 8001 --max-model-len 16000

python -m starvector.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port <different from 40000, say 40001> --worker http://localhost:<change accordingly, i.e. 40001> --model-path <ckpt2>

Option 2: Launch VLLM

0. Remember to clone the starvector/vllm fork (it has modifications for starvector).

git clone https://github.com/starvector/vllm.git
cd vllm
pip install -e .

1. Call this to launch the VLLM endpoint

vllm serve starvector/starvector-1b-im2svg --chat-template configs/chat-template.jinja --trust-remote-code --port 8000 --max-model-len 8192

2. Create the demo for VLLM

python -m starvector.serve.vllm_api_gradio.controller --host 0.0.0.0 --port 10000
python -m starvector.serve.vllm_api_gradio.gradio_web_server --controller http://localhost:10000 --model-list-mode reload --port 7000
python -m starvector.serve.vllm_api_gradio.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-name starvector/starvector-1b-im2svg --vllm-base-url http://localhost:8000

3. Add more models by serving them with VLLM and calling a new model worker

vllm serve starvector/starvector-8b-im2svg --chat-template configs/chat-template.jinja --trust-remote-code --port 8001 --max-model-len 16384

python -m starvector.serve.vllm_api_gradio.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40001 --worker http://localhost:40001 --model-name starvector/starvector-8b-im2svg --vllm-base-url http://localhost:8001

Citation

@misc{rodriguez2024starvector,
      title={StarVector: Generating Scalable Vector Graphics Code from Images and Text}, 
      author={Juan A. Rodriguez and Abhay Puri and Shubham Agarwal and Issam H. Laradji and Pau Rodriguez and Sai Rajeswar and David Vazquez and Christopher Pal and Marco Pedersoli},
      year={2024},
      eprint={2312.11556},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2312.11556}, 
}

License

This project is licensed under the Apache License, Version 2.0 - see the LICENSE file for details.