---
license: mit
language:
- en
base_model:
- deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
tags:
- svg
- text-to-svg
- vector-graphics
- code-generation
- reasoning
- qwen2
- safetensors
- arxiv:2509.24299
widget:
- text: "Generate an SVG icon of a yellow rubber duck facing right on a shallow blue oval of water."
- text: "Generate a minimalist calendar icon with a checkmark below it."
- text: "Generate a flat vector icon of a delivery truck using bold solid colors."
---

# SVGThinker-7B

SVGThinker-7B is a text-to-SVG generation model introduced in
[SVGThinker: Instruction-Aligned and Reasoning-Driven Text-to-SVG Generation](https://arxiv.org/abs/2509.24299).
It generates editable SVG code from natural-language descriptions, with a focus on
compact icon-style vector graphics.

The model is fine-tuned from
[`deepseek-ai/DeepSeek-R1-Distill-Qwen-7B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)
and is released as BF16 sharded safetensors.

## Links

- Paper: [arXiv:2509.24299](https://arxiv.org/abs/2509.24299)
- Demo Space: [Caffin/SVGThinker-7B](https://huggingface.co/spaces/Caffin/SVGThinker-7B)
- License: MIT

## Intended Use

This model is intended for:

- generating SVG icons from English text prompts
- prototyping simple vector graphics
- producing editable SVG markup rather than raster images
- research on text-to-SVG and structured code generation

Generated SVG should be reviewed and sanitized before being rendered in production
web pages or downstream applications.

## Quick Start

```bash
pip install "transformers>=4.51.0" torch accelerate safetensors
```

```python
import re
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Caffin/SVGThinker-7B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    use_safetensors=True,
)

description = "A minimalist calendar icon with two black tabs and a bold checkmark below it."
prompt = "Review the given information below and generate a svg according to it.\n" + description

messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    output = model.generate(
        inputs,
        max_new_tokens=4096,
        do_sample=True,
        temperature=0.8,
        top_p=0.6,
        repetition_penalty=1.05,
    )

text = tokenizer.decode(output[0], skip_special_tokens=False)
match = re.search(r"<svg.*?</svg>", text, flags=re.DOTALL)
print(match.group(0) if match else text)
```

The model may emit reasoning text before the final SVG. For most applications,
extract the `<svg>...</svg>` block before rendering.

## Model Notes

SVGThinker is trained directly in SVG code space. The paper describes a sequential
annotation pipeline that aligns natural-language descriptions with the step-by-step
construction of SVG primitives, helping the model generate more editable SVG code.

For full training data, annotation, and evaluation details, see the paper.

## Evaluation Snapshot

On the paper's 1,000 held-out text-to-SVG prompts, SVGThinker reports:

| Model | FID lower is better | CLIP higher is better | FID-CLIP lower is better | Primitive support |
| --- | ---: | ---: | ---: | --- |
| SVGThinker-7B | 34.06 | 0.2765 | 21.08 | all |

## Limitations

- Outputs may be malformed, incomplete, or visually inconsistent with the prompt.
- The model is best suited for simple to moderately complex icon-style graphics.
- It may struggle with photorealistic scenes, dense layouts, and text-heavy SVGs.
- SVG is executable markup in browser contexts; treat generated SVG as untrusted.
- The model primarily targets English prompts.

## Citation

```bibtex
@inproceedings{chen2025svgthinker,
  title = {SVGThinker: Instruction-Aligned and Reasoning-Driven Text-to-SVG Generation},
  author = {Chen, Hanqi and Zhao, Zhongyin and Chen, Ye and Liang, Zhujin and Ni, Bingbing},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
  year = {2025},
  publisher = {ACM},
  doi = {10.1145/3746027.3755392},
  url = {https://arxiv.org/abs/2509.24299}
}
```