HiVG: Hierarchical SVG Tokenization

This repository contains the official implementation of the paper "Hierarchical SVG Tokenization: Learning Compact Visual Programs for Scalable Vector Graphics Modeling".

Highlights

• Small Model, Frontier Results — 3B parameters that beat 7/7 proprietary models including GPT-5 and Gemini 2.5 on image-to-SVG.
• Efficient SVG Token Compression — Hierarchical tokenization (Raw SVG → Atomic tokens → Segment tokens) with 2.76x sequence compression.
• High-Fidelity Image-to-SVG — Convert any image into a clean, editable SVG — structure, layout, and detail faithfully preserved.

Table of Contents

• Installation
• Quick Start
• Model Weights
• Evaluation
• Segment Token Training
• Generation Parameters
• Citation

Installation

git clone https://github.com/ximinng/HiVG.git
cd HiVG
pip install -e .

Quick Start

Python API

from hivg_infer import HiSVGInferencePipeline

pipeline = HiSVGInferencePipeline(
    model_path="/path/to/model",
    coord_range=234,
    temperature=0.7,
    top_p=0.9,
    max_new_tokens=4096,
)

# Image-to-SVG
result = pipeline.img2svg("assets/cases/w2.png")
if result["success"]:
    print(result["svg"])

# Text-to-SVG
result = pipeline.text2svg("The image features a simple, flat-style briefcase with a handle and a central plus symbol. The briefcase is outlined in black with a light teal plus sign, suggesting medical or emergency services. Small decorative dots surround the briefcase, enhancing its minimalistic design.")

if result["success"]:
    with open("output.svg", "w") as f:
        f.write(result["svg"])

Command Line

# Text-to-SVG
python -m hivg_infer.cli --model_path /path/to/model \
    --prompt "A minimalist black phone icon with an outline style"

# Image-to-SVG
python -m hivg_infer.cli --model_path /path/to/model \
    --image photo.png

# Batch inference
python -m hivg_infer.cli --model_path /path/to/model \
    --dataset test.json --output_dir ./outputs

# Interactive mode
python -m hivg_infer.cli --model_path /path/to/model --interactive

Model Weights

Evaluation

# HuggingFace backend
python -m hivg_metric.eval \
    --model_path /path/to/model \
    --dataset test.json \
    --format alpaca \
    --coord_range 234 \
    --save_name results.jsonl \
    --save_svg_dir ./svgs/

# vLLM backend (faster)
python -m hivg_metric.eval \
    --model_path /path/to/model \
    --dataset test.json \
    --infer_backend vllm \
    --batch_size 32 \
    --save_name results.jsonl

# Text-to-SVG: CLIP score + preference metrics
python -m hivg_metric.compute_metrics \
    --input results.jsonl \
    --task_type text2svg \
    --metrics basic,clip,preference \
    --output metrics.json

# Image-to-SVG: visual similarity metrics
python -m hivg_metric.compute_metrics \
    --input results.jsonl \
    --task_type img2svg \
    --image_base_dir /path/to/images \
    --metrics basic,visual \
    --output metrics.json

python -m hivg_metric.visualize \
    --input results.jsonl \
    --output report.html \
    --title "HiVG Evaluation Report"

Supported Metrics

Segment Token Training

HiVG's key innovation is learning structure segment tokens that compress SVG command sequences. See the full guide:

Segment Token Training Guide — Step-by-step instructions for training BPE segments from your own SVG corpus and converting datasets.

Quick overview of the three-level tokenization:

Level 0 (Raw SVG)    path d="M 100 200 c -5 0 -10 5 -10 10"
                                    ↓
Level 1 (Atomic)     <cmd_M><P_100><P_200><cmd_c><d_-5><d_0><d_-10><d_5><d_-10><d_10>
                                    ↓
Level 2 (Segment)    <cmd_M><P_100><P_200><SEG_42>

Generation Parameters

Dataset Format

[
  {"instruction": "Draw a red circle", "input": "", "output": "<svg tokens>"}
]

[
  {
    "messages": [{"role": "user", "content": "<image>Convert to SVG"}],
    "images": ["path/to/image.png"],
    "label": "<svg tokens>"
  }
]

License

This project is licensed under the MIT License.

Citation

If you find HiVG useful in your research, please consider citing:

@article{xing2026hivg,
    title={Hierarchical SVG Tokenization: Learning Compact Visual Programs for Scalable Vector Graphics Modeling},
    author={Xing, Ximing and Xue, Ziteng and Li, Zhenxi and Liang, Weicong and Wang, Linqing and Yang, Zhantao and Hang, Tiankai and Yin, Zijin and Lu, Qinglin and Wang, Chunyu and Yu, Qian},
    journal={arXiv preprint arXiv:2604.05072},
    year={2026}
}