StarVector
ServiceNow Research, Mila, ETS Montreal · Image · Coding
StarVector is an open-source vision-language model that turns images or text into SVG code. Built by ServiceNow Research with Mila and ETS Montreal, it ships in 1B and 8B versions.
About StarVector
What Is StarVector
StarVector generates Scalable Vector Graphics by writing SVG code instead of tracing curves. Traditional vectorizers look at pixel edges and guess where the lines should go. StarVector reads an image first, understands what it shows, then writes the markup that rebuilds it with shapes like ellipses, polygons, and text. The same model also takes a written prompt and produces a brand new SVG from scratch.
The project comes from ServiceNow Research with Mila and ETS Montreal, and the code sits on GitHub under an Apache 2.0 license. You don't rent it as a service. You download the weights from Hugging Face and run them on your own hardware. That makes it a developer tool rather than a consumer app, so expect a Python setup and a GPU.
Two variants exist. The 8B model scores highest on the team's own benchmarks, while the 1B version trades some accuracy for lighter hardware needs. Both were trained on SVG-Stack, a dataset of about 2 million samples.
The biggest limit is what it was trained to see. The team states plainly that StarVector struggles with natural photos and detailed illustrations. It does best on icons, logos, technical diagrams, graphs, and charts. Feed it a portrait and the output will disappoint.
Getting Started
- Clone the star-vector repository from GitHub and create a Python 3.11 environment.
- Install the package with pip and pull the model weights you want, either the 1B or 8B checkpoint from Hugging Face.
- Load the model in Python and move it onto a CUDA GPU.
- Pass an image for vectorization or a text prompt for a fresh SVG.
- Rasterize the returned SVG code to check the result, then export the file.
Product Information
A quick look at StarVector's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Developers who need vector graphics at scale
- Design engineers wiring up a pipeline
- Researchers and students
- Teams that cannot send assets to a cloud service
Tasks
- Icons and logos into SVG
- Technical diagrams and charts
- Text-to-SVG on demand
- Batch re-vectorizing a library
Scenarios
- Rebuilding a stale icon set
- Turning a scanned chart into an editable graphic
- Prototyping a logo from a brief
- Keeping sensitive diagrams in house
Key features
Image-to-SVG vectorization
StarVector takes a raster image and writes the SVG code that recreates it. It understands what the picture shows and picks the right primitives, such as ellipses, polygons, and text, rather than dumping thousands of path points. The result is a compact file you can open and edit in any vector tool.
Text-to-SVG generation
Give the model a written description and it produces a fresh SVG with no source image at all. This is the part no classic vectorizer can do, since tracing tools need something to trace. It suits quick drafts of simple shapes and icons.
Semantic use of SVG primitives
The model chooses between shapes instead of falling back on paths for everything. A circle becomes a circle element and a label becomes a text element. That makes the output cleaner and easier for other tools to read and restyle later.
Two model sizes
StarVector ships in 1B and 8B versions with the same task setup. The 8B leads on accuracy, while the 1B runs on lighter hardware. You pick the trade-off between quality and resources for your own setup.
SVG-Bench and SVG-Stack
The team released a benchmark across 10 datasets and three tasks, plus a training set of roughly 2 million samples. Both are public on Hugging Face, so you can reproduce the numbers or fine-tune on your own data.
Open weights and Apache 2.0 license
The code and model weights are free to download and use, including commercially. There is no hosted API and no usage bill, so the only cost is the machine you run it on.
Pros and cons
Pros
- Writes real SVG shapes instead of dense paths, so files stay compact and easy to edit.
- Open weights under Apache 2.0 let you run it locally and even ship it in a product.
- Handles both image input and text prompts, covering vectorization and fresh generation.
- Strong on icons, logos, and diagrams, with published scores on a public benchmark.
- The 1B model gives a lighter option when an 8B checkpoint is too heavy for your GPU.
Cons
- Performs poorly on natural photos and detailed illustrations, which the team openly states, so you cannot point it at any image and expect a good trace.
- There is no hosted API or one-click app, meaning setup takes Python and a CUDA GPU, which rules out non-technical users.
- Output quality varies by image type, so batches of mixed content still need a human check.
Frequently asked questions
It converts images into SVG code and generates SVG from text. Typical uses are turning icons, logos, and diagrams into clean, editable vector files, or producing simple shapes from a written prompt.
Related content
Explore related tools, skills, and articles for StarVector.
StarVector Alternatives
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
Tractionwave
TractionWave · Image · MarketingTractionWave is an AI attention heatmap tool for static ads. Upload a PNG or JPG and it returns predicted attention, ranked hotspots, and fixes before you spend budget.
Imejis.io
Imejis.io · Image · MarketingImejis.io is an AI design studio for agents and marketing teams. Describe an image in plain words, and it writes the copy, builds the layout, reflows it to every size, then exports a finished file.